Engine 0.1.21: one model across two or three GPUs (layer split, experimental); setup requires it

Fitted on the 5080 + 3090 K sweep (18-32): it picks K=26, the measured best (prompt +18-20%
over the 5080 alone), and gives a slow third card (2080 Ti) a single layer.

.gitignore

@@ -46,3 +46,4 @@ build-vision.bat
 /dist/
 *.done
 /third_party/*.zip
+

CMakeLists.txt

@@ -8,7 +8,14 @@
 # written against sm_120 (RTX 5000 / Blackwell) and silently building for an older arch would produce a
 # binary that runs and is wrong about its own performance, which is the worst of both outcomes.
 cmake_minimum_required(VERSION 3.24)
-project(strata VERSION 0.1.0 LANGUAGES CXX)
+project(strata VERSION 0.1.21 LANGUAGES CXX)
+add_compile_definitions(STRATA_VERSION="${PROJECT_VERSION}")
+# issue #31: release builds keep a PDB beside the exe (not shipped in it), so a minidump from a stalled engine can
+# be read back to the exact source line
+if(MSVC AND CMAKE_BUILD_TYPE STREQUAL "Release")
+  add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/Zi>)
+  add_link_options(/DEBUG /OPT:REF /OPT:ICF)
+endif()
 
 # The prebuilt release (setup.py downloads it, so users need no compiler): ggml's CPU backend for any AVX2 CPU
 # instead of this machine's (Strata's own AVX-512 kernels are per-file and chosen at run time either way), and
@@ -49,7 +56,7 @@ if(STRATA_ENABLE_CUDA)
     if(_base STREQUAL "native" OR _base STREQUAL "all" OR _base STREQUAL "all-major")
       continue()
     endif()
-    if(_base LESS 80)
+    if(_base LESS 80 AND NOT STRATA_EXPERIMENTAL_SM75)   # the A/B of a Turing card as a layer-split stage
       message(FATAL_ERROR
         "Strata needs an NVIDIA GPU of compute capability 8.0 or newer (RTX 30 / 40 / 50 series), but "
         "CMAKE_CUDA_ARCHITECTURES is '${CMAKE_CUDA_ARCHITECTURES}'.")
@@ -177,7 +184,10 @@ if(STRATA_ENABLE_CUDA)
                           src/kernels/cuda/native_qsa_score.cu
                           src/kernels/cuda/ple.cu
                           src/kernels/cuda/kv_q8.cu
+                          src/kernels/cuda/kv_stream.cu
+                          src/kernels/cuda/kv_q4.cu
                           src/kernels/cuda/fused_gr.cu
+                          src/kernels/cuda/cvec.cu
                           src/kernels/cuda/qsa_decode_attn.cu
                           src/kernels/cuda/fused_gdn.cu
                           src/kernels/cuda/dequant_bf16.cu
@@ -235,7 +245,7 @@ if(STRATA_ENABLE_CUDA)
   # already links `strata_core` - putting `gdn_layer` in core would make the two mutually dependent.  A third
   # library is the honest shape of the dependency graph: core knows about devices and weights, kernels know
   # about arithmetic, and this one knows the ORDER the arithmetic happens in.
-  add_library(strata_engine STATIC src/core/layer.cpp src/core/session.cpp src/core/expert_source.cpp
+  add_library(strata_engine STATIC src/core/layer.cpp src/core/session.cpp src/core/expert_source.cpp src/core/remote_experts.cpp
     src/core/expert_cache.cpp src/core/native_head.cpp src/core/native_dense.cpp src/core/verify.cpp src/core/mtp.cpp)
   target_include_directories(strata_engine PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
   target_link_libraries(strata_engine PUBLIC strata_core strata_kernels strata_kernels_cpu CUDA::cudart)
@@ -250,7 +260,7 @@ if(STRATA_ENABLE_CUDA)
 
   add_executable(strata src/program/generate.cpp)
   target_include_directories(strata PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(strata PRIVATE strata_engine strata_prefill CUDA::cudart)
+  target_link_libraries(strata PRIVATE strata_engine strata_prefill strata_spec CUDA::cudart)
 
   add_executable(dequant_s2_parity src/kernels/dequant_s2_parity.cpp)
   # The tests are HOST C++ files that call cudaMalloc/cudaMemcpy themselves, so they need the toolkit headers
@@ -356,6 +366,22 @@ if(STRATA_ENABLE_CUDA)
   target_link_libraries(kv_q8_parity PRIVATE strata_kernels CUDA::cudart)
   add_test(NAME kv_q8_parity COMMAND kv_q8_parity)
 
+  # KV streaming (docs/kv-streaming-design.md): streamed vs resident attention, bitwise (GPU, synthetic, no model).
+  add_executable(kv_stream_parity src/kernels/kv_stream_parity.cpp)
+  target_include_directories(kv_stream_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  target_link_libraries(kv_stream_parity PRIVATE strata_kernels CUDA::cudart)
+  add_test(NAME kv_stream_parity COMMAND kv_stream_parity)
+
+  # PR #21: Q4_0 KV with Walsh-Hadamard rotation vs a host reference (GPU, synthetic, no model).
+  add_executable(kv_q4_parity src/kernels/kv_q4_parity.cpp)
+  target_include_directories(kv_q4_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  target_link_libraries(kv_q4_parity PRIVATE strata_kernels CUDA::cudart)
+  add_test(NAME kv_q4_parity COMMAND kv_q4_parity)
+  add_executable(cvec_parity src/kernels/cvec_parity.cpp)
+  target_include_directories(cvec_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+  target_link_libraries(cvec_parity PRIVATE strata_kernels CUDA::cudart)
+  add_test(NAME cvec_parity COMMAND cvec_parity)
+
   # Plan v0.3 P3: native MMVQ for 1..8 tokens (bitwise selftest + --bench timing); synthetic data, no model.
   if(STRATA_BUILD_TESTS)
   add_executable(native_mmvq_multi bench/micro/native_mmvq_multi.cpp)
@@ -417,7 +443,7 @@ endif()
 #     worse opt-out.  `cpu_require_expert_support()` therefore refuses at startup on a CPU without the
 #     features, and the fallback exists for tests, exactly as P2.S3 specifies.
 # Plan v0.3 P6: speculation components that need no GPU.
-add_library(strata_spec STATIC src/spec/suffix_drafter.cpp src/spec/controller.cpp)
+add_library(strata_spec STATIC src/spec/suffix_drafter.cpp src/spec/controller.cpp src/spec/draft_policy.cpp)
 target_include_directories(strata_spec PUBLIC include)
 if(STRATA_BUILD_TESTS)
   add_executable(suffix_drafter_test src/spec/suffix_drafter_test.cpp)
@@ -426,6 +452,13 @@ if(STRATA_BUILD_TESTS)
   add_executable(controller_test src/spec/controller_test.cpp)
   target_link_libraries(controller_test PRIVATE strata_spec)
   add_test(NAME controller_test COMMAND controller_test)
+  add_executable(draft_policy_test src/spec/draft_policy_test.cpp)
+  target_link_libraries(draft_policy_test PRIVATE strata_spec)
+  add_test(NAME draft_policy_test COMMAND draft_policy_test)
+  # the serve loop's conversation-cache retention policy (header-only, no library to link)
+  add_executable(conv_cache_test src/program/conv_cache_test.cpp)
+  target_include_directories(conv_cache_test PRIVATE include)
+  add_test(NAME conv_cache_test COMMAND conv_cache_test)
 endif()
 
 add_library(strata_kernels_cpu STATIC src/kernels/cpu/expert.cpp src/kernels/cpu/pool.cpp
@@ -485,16 +518,43 @@ if(STRATA_NATIVE_EXPERTS)
   endif()
   # part of strata_kernels_cpu (the pool calls these and they call its CPU probe: one library, so GNU ld's
   # one-pass static linking has no cycle to resolve)
-  target_sources(strata_kernels_cpu PRIVATE src/kernels/cpu/native_expert.cpp src/kernels/cpu/iq_avx512.cpp)
+  target_sources(strata_kernels_cpu PRIVATE src/kernels/cpu/native_expert.cpp src/kernels/cpu/iq_avx512.cpp
+                                           src/kernels/cpu/iq_avx2.cpp)
   target_include_directories(strata_kernels_cpu PRIVATE third_party/ggml)
   if(MSVC)
     set_source_files_properties(src/kernels/cpu/iq_avx512.cpp PROPERTIES COMPILE_OPTIONS "/arch:AVX512")
+    set_source_files_properties(src/kernels/cpu/iq_avx2.cpp PROPERTIES COMPILE_OPTIONS "/arch:AVX2")
   else()
     set_source_files_properties(src/kernels/cpu/iq_avx512.cpp PROPERTIES COMPILE_OPTIONS
                                 "-mavx512f;-mavx512bw;-mavx512vl;-mavx512dq;-mavx512vnni;-mavx512vbmi;-mf16c;-mfma")
+    set_source_files_properties(src/kernels/cpu/iq_avx2.cpp PROPERTIES COMPILE_OPTIONS "-mavx2;-mfma;-mf16c")
   endif()
   target_link_libraries(strata_kernels_cpu PUBLIC ggml-cpu ggml-base)
   target_compile_definitions(strata_kernels_cpu PUBLIC STRATA_NATIVE_EXPERTS=1)
+  if(STRATA_ENABLE_CUDA)
+    # prompt-speed plan step 2b: llama.cpp's MMQ kernels (ggml-cuda, MIT) for the prompt path's experts, compiled
+    # from the same checkout as ggml above - only mmq.cuh's instances for the packs' types and the quantizer, with
+    # the few host symbols of ggml-cuda.cu they use in src/prefill/ggml_cuda_host.cu
+    if(STRATA_GGML_DIR)
+      set(_strata_gg ${STRATA_GGML_DIR}/ggml)
+    else()
+      set(_strata_gg ${strata_llamacpp_SOURCE_DIR}/ggml)
+    endif()
+    set(_strata_mmq ${_strata_gg}/src/ggml-cuda)
+    add_library(strata_mmq STATIC src/prefill/moe_mmq.cu src/prefill/ggml_cuda_host.cu ${_strata_mmq}/quantize.cu)
+    foreach(_t q2_0 iq2_xxs iq2_xs iq2_s iq3_xxs iq3_s iq4_nl iq4_xs)
+      target_sources(strata_mmq PRIVATE ${_strata_mmq}/template-instances/mmq-instance-${_t}.cu)
+    endforeach()
+    target_include_directories(strata_mmq PRIVATE ${_strata_gg}/include ${_strata_gg}/src ${_strata_mmq}
+                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    target_compile_options(strata_mmq PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda -use_fast_math>)
+    if(MSVC)
+      target_compile_options(strata_mmq PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/bigobj>)
+    endif()
+    target_link_libraries(strata_mmq PUBLIC ggml-base CUDA::cudart)
+    target_link_libraries(strata_prefill PUBLIC strata_mmq)
+    target_compile_definitions(strata_prefill PUBLIC STRATA_PREFILL_MMQ=1)
+  endif()
   if(STRATA_ENABLE_CUDA)
     add_executable(native_expert_parity src/kernels/native_expert_parity.cpp)
     target_include_directories(native_expert_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
@@ -528,9 +588,16 @@ if(STRATA_BUILD_TESTS)
   add_executable(pool_test src/kernels/cpu/pool_test.cpp)
   target_link_libraries(pool_test PRIVATE strata_kernels_cpu)
   add_test(NAME pool_test COMMAND pool_test --selftest WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+  # issue #29: tiny batches with sleeping workers, as on a big-VRAM card; a hang fails it
+  add_executable(pool_stress src/kernels/cpu/pool_stress.cpp)
+  target_link_libraries(pool_stress PRIVATE strata_kernels_cpu)
+  add_test(NAME pool_stress COMMAND pool_stress 20)
 endif()
 
 if(STRATA_BUILD_TESTS)
   enable_testing()
-  add_subdirectory(tests)
+  # the unit tests above build from src/; the tests/ tree (model-level checks) is not part of the published repo
+  if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/tests/CMakeLists.txt)
+    add_subdirectory(tests)
+  endif()
 endif()

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Niko1221 and the Strata contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -3,131 +3,197 @@
 <p align="center"><b>Run a 125-billion-parameter AI model on a normal gaming PC</b><br>
 one NVIDIA card (12-24 GB) + 64 GB of RAM · Windows or Linux · one click to install</p>
 
+<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
+<sub>A voxel pagoda garden, 1 shot prompt running on an RTX 5070 with Strata (IQ3_S, 128K context) ·
+<a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4">full video (49 s)</a></sub></p>
+
 Strata runs **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)** - a large, smart AI model that
 normally needs a server - on your own PC. It writes its answers at **60-95 tokens per second** (a token is about ¾
 of a word): faster than you can read.
 
-- **Private:** everything runs on your PC. Nothing is sent anywhere.
-- **Works with your apps:** chat apps, coding agents and scripts that speak the OpenAI or Anthropic API just work.
-- **Sees pictures** too, if you want (screenshots, photos, scanned pages).
 - **Free and open source.**
 
-> **Jump to:** [Is my PC enough?](#is-my-pc-enough) · [Install](#install-3-steps) · [How fast?](#how-fast-is-it) ·
-> [Which model?](#which-model-should-i-pick) · [Using it](#using-it) · [Problems?](#something-went-wrong) ·
-> [How it works](#how-does-it-work) · [All the details](docs/DETAILS.md)
+> **Jump to:** [How fast?](#how-fast-is-it) · [Which model?](#which-model-should-i-pick) · [Install](#install) ·
+> [Using it](#using-it) · [Problems?](#something-went-wrong) · [How it works](#how-does-it-work) ·
+> [All the details](docs/DETAILS.md)
 
 ---
 
-## Is my PC enough?
-
-| You need | |
-| --- | --- |
-| **Graphics card** | NVIDIA RTX 30, 40 or 50 series with **12 GB of VRAM or more** |
-| **Memory (RAM)** | **64 GB** |
-| **Free disk space** | ~80 GB (an SSD makes the first start much faster) |
-| **System** | Windows 10/11, or Linux |
-
-That's it. The only thing you install yourself is a current **NVIDIA driver**
-([nvidia.com/drivers](https://www.nvidia.com/drivers) or the NVIDIA App). Everything else - Python, the engine, the
-model - is set up for you.
-
-## Install (3 steps)
-
-**Windows**
-
-1. [Download this project](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) and unzip it (or `git clone` it).
-2. Double-click **`START-HERE.bat`**.
-3. Answer 4 questions - or just press Enter each time for the recommended choice:
-   - **Which model?** The original, or Swift 1.5 (a version that thinks shorter and answers sooner)
-   - **Which size?** Q2_0, IQ2_XS or IQ3_XXS - see [which model](#which-model-should-i-pick)
-   - **How much context?** How much text it can keep in mind at once (it suggests one for your card)
-   - **Images?** Whether it should also read pictures
-
-Then it downloads everything (the model is ~70 GB, so the first time takes a while - you can stop and it picks up
-where it left off) and **starts the model**. Your browser opens a chat page at `http://127.0.0.1:8080`.
-
-**Next time**, just double-click `START-HERE.bat` again: it starts right away, nothing is downloaded twice. Close its
-window to stop the model.
-
-**Linux:** run `./setup.sh` - same questions, same result.
-
 ## How fast is it?
 
 Measured on an RTX 5070 (12 GB), a Ryzen 5 7600 and 64 GB of RAM:
 
 | Size | Writes answers (short chat) | Writes answers (128K context) | Reads your prompt |
 | --- | ---: | ---: | ---: |
-| **Q2_0** | 95 tokens/s | 65 tokens/s | 539 tokens/s |
-| **IQ2_XS** | 78 tokens/s | 52 tokens/s | 463 tokens/s |
-| **IQ3_XXS** | 66 tokens/s | 45 tokens/s | 410 tokens/s |
+| **Q2_0** | 90 tokens/s | 67 tokens/s | 1,310 tokens/s |
+| **IQ2_XS** | 74 tokens/s | 60 tokens/s | 1,240 tokens/s |
+| **IQ3_XXS** | 62 tokens/s | 46 tokens/s | 1,110 tokens/s |
+| **IQ3_S** | 52 tokens/s | 41 tokens/s | 1,070 tokens/s |
+| **Coder** (IQ1_M) | 51 tokens/s | 44 tokens/s | 1,300 tokens/s |
 
 - **Writes answers** = how fast the reply appears (tokens per second).
-- **Reads your prompt** = how fast it takes in what you send (long documents, code, chat history).
+- **Reads your prompt** = how fast it takes in what you send (long documents, code, chat history), measured on a
+  32K-token prompt; a 4K prompt reads at 740-1,000 tokens/s. A 32K prompt takes about 25 seconds with Q2_0.
 
 A card with more VRAM is faster, because more of the model fits on the GPU: an RTX 3090 (24 GB) should do roughly
 100-140 tokens per second. All measurements, long-context numbers and estimates for other cards are in the
 [details](docs/DETAILS.md#speed-measured).
 
+Every PC is different: `START-HERE.bat --calibrate` measures a few engine settings on yours and keeps the fastest
+(about 5-10 minutes; on the PC above it made the Coder 7% faster).
+
+**Two or three NVIDIA cards?** `START-HERE.bat --setup --gpus 0,2` splits the model's layers across them (experimental):
+each card keeps the experts of its own layers, and prompts flow through the cards in a pipeline. On an RTX 5080 +
+RTX 3090 it read prompts 18-20% faster than the 5080 alone, with decoding on par. See [docs/MULTI_GPU.md](docs/MULTI_GPU.md).
+
 ## Which model should I pick?
 
 **The size** (the same model, compressed more or less):
 
-| Size | Download | Speed | Quality | Pick it if... |
-| --- | ---: | --- | --- | --- |
-| **Q2_0** | 66 GB | fastest | good | you want speed |
-| **IQ2_XS** | 68 GB | fast | better | you want a good all-rounder (**recommended**) |
-| **IQ3_XXS** | 76 GB | slower | best | you want the best answers (uses 43 GB of your 64 GB RAM) |
+| Model | RAM+VRAM Requirements | Speed | Quality |
+| --- | ---: | --- | --- |
+| **Q2_0** | 37.6 GB | fastest | good |
+| **IQ2_XS** | 39.2 GB | fast | better (**recommended**) |
+| **IQ3_XXS** | 47.0 GB | slower | great |
+| **IQ3_S** | 54.8 GB | slowest | best: matches the full model on the published tests (original model only) |
+
+**Will it fit?** Shard 1 is the part of the model that gets loaded when it starts: its experts go into your **RAM**,
+the rest onto your graphics card (the second shard, a 29 GB lookup table, stays on the SSD). So it fits when your
+**RAM is at least shard 1 + about 10 GB** for Windows and your other programs. With 64 GB of RAM every size fits
+(IQ3_S with little else open); with 48 GB, Q2_0 and IQ2_XS. A bigger graphics card makes it faster, but it doesn't
+lower the RAM needed.
 
 **The version:**
 
 - **Qwen3.8-Flash-Next** - the original.
+- **[Coder](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF)** - ISTA-DASLab's coding
+  version: half of the experts removed, keeping the ones that code, tool use and images need (91% of the full model's
+  SWE-bench Verified score, 99% of LiveCodeBench, by its authors). One size (IQ1_M: its experts stored like IQ3_S):
+  shard 1 is **29.6 GB**, so it fits a PC with **32 GB of RAM**, runs 262K context on 64 GB, and reads long prompts
+  the fastest of all. Weaker outside coding.
 - **[Swift 1.5](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF)** - a fine-tune by UkisAI
   that thinks much shorter before answering, so you get the answer sooner, with about the same quality. Same speed per
-  token. Its own license applies (see its page).
+  token, and about the same RAM as the same size of the original (no IQ3_S). Its own license applies (see its page).
 
-Not sure? Take **IQ2_XS**. You can add another one later with `START-HERE.bat --setup`.
+Not sure? Take **IQ2_XS** - or the **Coder** if you mainly write code, or have 32-48 GB of RAM. You can add another
+one later with `SETUP.bat` (the same as `START-HERE.bat --setup`; on Linux `./setup.sh --setup`).
+
+For **OrcaRouter's Flash-Next Uncensored IQ3_XXS**, see the [manual compatibility setup](docs/ORCA.md).
+It needs an explicit packing conversion and is not an installer menu option.
+
+## Install
+
+**You need:** an NVIDIA RTX 30, 40 or 50 card with 12 GB of VRAM or more, enough RAM for the size you pick (above),
+~80 GB of free disk space (an SSD makes the first start much faster), and Windows 10/11 or Linux. The only thing you
+install yourself is a current **NVIDIA driver** ([nvidia.com/drivers](https://www.nvidia.com/drivers) or the NVIDIA
+App). Everything else - Python, the engine, the model - is set up for you.
+
+**Windows**
+
+1. [Download this project](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) and unzip it (or `git clone` it).
+2. Double-click **`START-HERE.bat`**.
+3. Answer a few questions - or just press Enter each time for the recommended choice:
+   - **Which model and size?** The original or Swift 1.5, and Q2_0, IQ2_XS, IQ3_XXS or IQ3_S - see [above](#which-model-should-i-pick)
+   - **How much context?** How much text it can keep in mind at once (it suggests one for your card)
+   - **Images?** Whether it should also read pictures
+   - **Experimental speed projection?** Off unless you say yes - [read what it does](docs/DETAILS.md#experimental-speed-projection-experimental-off-by-default) first
+
+Then it downloads everything (the model is ~70 GB, so the first time takes a while - you can stop and it picks up
+where it left off) and **starts the model**. Your browser opens the Strata app at `http://127.0.0.1:8080`.
+
+> **While the model starts, your PC can be slow or stop responding for 1-3 minutes** (longest the first time): Strata
+> loads 35-55 GB into your RAM and locks part of it for the graphics card. That's normal - wait, and don't close the
+> window. The window tells you what it is doing.
+
+**Next time**, just double-click `START-HERE.bat` again: it starts right away, nothing is downloaded twice. Close its
+window to stop the model.
+
+**Updating:** download the new version and unzip it anywhere (or `git pull`), then run `START-HERE.bat` in it. The
+model files are kept in a `Strata-data` folder next to your Strata folder, so a new copy finds them and sets itself up
+the same way - nothing big is downloaded again.
+
+**Linux:** run `./setup.sh` - same questions, same result.
 
 ## Using it
 
-- **Chat in the browser:** `http://127.0.0.1:8080` - a simple chat page (it opens by itself when the model starts).
+<p align="center"><img src="docs/media/runpagoda.png" width="900" alt="The Strata app's Monitor tab next to a coding agent"><br>
+<sub>The Strata app's <b>Monitor</b> (left) while a coding agent writes the pagoda garden from the video (right)</sub></p>
+
+- **In the browser:** `http://127.0.0.1:8080` - the Strata app (it opens by itself when the model starts): **Chat**, a
+  live **Monitor** of the model and your GPU/CPU/RAM, and **About** with the settings and addresses.
 - **Chat in the terminal:** `.venv\Scripts\python chat.py`
 - **Your apps and coding agents:** add it as an "OpenAI-compatible" provider with base URL
   **`http://127.0.0.1:8080/v1`**, any API key and any model name. Apps that use Anthropic's API: `http://127.0.0.1:8080/v1/messages`.
 - **Thinking:** the model thinks before it answers. Choose **off, low, medium or high** - in the chat page menu, with
   `/think low` in `chat.py`, or with your app's "reasoning effort" setting. Off is fastest; high is best for hard questions.
 - **Pictures:** in the chat page click **Picture**; in `chat.py` type `/image <path>`; in apps just attach them.
-- **From your phone or another PC:** see the [details](docs/DETAILS.md#using-it) (set an API key first).
+- **From your phone or another PC:** `START-HERE.bat --setup --host 0.0.0.0 --api-key <secret>`, then open the
+  address the server window prints; see the [details](docs/DETAILS.md#using-it).
+- **Experimental speed projection (off by default):** an experimental control vector that setup can turn on; it
+  changes how the model answers - read [what it does](docs/DETAILS.md#experimental-speed-projection-experimental-off-by-default) first.
 
-**Good to know:** it answers one request at a time, and it re-reads the whole conversation for every answer. So in
-very long chats you wait longer before it starts writing: about 1 minute per 30,000 tokens of conversation.
+**Good to know:** it answers one request at a time. The first message of a chat is read in full (about 1 minute per
+30,000 tokens); after that it keeps the conversation and reads only what is new, so follow-ups start in seconds.
 
 ## Something went wrong?
 
-| What you see | What to do |
-| --- | --- |
-| `the NVIDIA driver is too old` | Update the driver (NVIDIA App or nvidia.com/drivers), restart the PC, run `START-HERE.bat` again. |
-| It stopped during download or setup | Run `START-HERE.bat` again - it continues where it stopped. |
-| `port 8080 is already in use` | Strata is already running - look for its window. |
-| The first start takes minutes | Normal: it loads 35-43 GB into RAM. The next start is faster. |
-| Slow, and the disk light is busy | Not enough free RAM: close other programs (browsers use a lot), or pick Q2_0 / IQ2_XS. |
-| "prompt exceeds the context" | The conversation is longer than the context you chose: run `START-HERE.bat --setup` and pick more. |
+**My PC froze, or got very slow, the first time Strata started.**
+That's normal while it starts, most of all the first time. Strata loads 35-55 GB into your RAM, locks part of it for
+the graphics card, and works out how much of the model fits on your GPU. The mouse can freeze for a few minutes. **Wait, and don't close the
+window.** The next starts are much faster. Still frozen after 10 minutes? Restart the PC, close other programs
+(browsers use a lot of RAM) and try again. If it keeps happening, pick a smaller size (Q2_0 or IQ2_XS).
 
-More in the [full troubleshooting table](docs/DETAILS.md#troubleshooting). Still stuck? Open an issue and attach
-`strata-<model>.log` from this folder.
+**It stopped while downloading or installing.**
+Run `START-HERE.bat` again. It continues where it stopped.
+
+**It says the NVIDIA driver is too old.**
+Update it (NVIDIA App or [nvidia.com/drivers](https://www.nvidia.com/drivers)), restart the PC, and run
+`START-HERE.bat` again.
+
+**It says port 8080 is already in use.**
+Strata is already running. Look for its window.
+
+**It's very slow and the disk light keeps blinking.**
+Your PC is out of free RAM. Close other programs, or pick a smaller size (Q2_0 or IQ2_XS).
+
+**An answer stopped with "the engine stopped unexpectedly".**
+Usually not enough RAM (on Linux the system then stops the engine). Just send your message again: Strata starts the
+engine by itself. If it keeps happening, close other programs or pick a smaller size.
+
+**It says the prompt exceeds the context.**
+The conversation is longer than the context you chose. Start a new chat, or run `SETUP.bat` and pick more
+context.
+
+**Still stuck?** Look in the [full troubleshooting table](docs/DETAILS.md#troubleshooting), or open an issue and
+attach `strata-<model>.log` from the Strata folder.
 
 ## How does it work?
 
-A model this big doesn't fit on a gaming graphics card. Strata splits the work between the parts of your PC:
+Models like this one normally run on servers with hundreds of gigabytes of graphics memory. Your graphics card has
+12-24 GB. Strata makes it fit by **sharing the work across your whole PC** - the same idea as a kitchen, where the
+things you use all the time stay on the counter and the rest waits in the pantry.
 
-<p align="center"><img src="docs/paper/tiers.svg" width="700" alt="how Strata splits the model between GPU, RAM and SSD"></p>
+<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
 
-- **The GPU** runs the part of the model that is used for every word, plus the "experts" it needs most often.
-- **The RAM** holds all 24,576 experts, and **the CPU** computes the few the GPU doesn't have - at the same time as the GPU.
-- **The SSD** holds a big lookup table; the model reads a few rows of it per word.
-- **A small helper inside the model guesses the next words**, and Strata checks several guesses at once. That makes
-  it 1.6-1.8x faster than going word by word - and the answer is exactly the same.
+- **The model is a team of 24,576 small specialists ("experts"),** and each word it writes needs only 10 of them.
+  So it doesn't have to have all of them on the graphics card at once.
+- **Your graphics card** does the part of the work needed for every word, and keeps the few thousand experts that
+  are asked most often. It keeps learning which ones those are while you use it.
+- **Your RAM** holds every expert. When a word needs one the card doesn't have, **your processor** works on it -
+  at the same time as the graphics card, so neither waits for the other.
+- **Your SSD** holds a big lookup table; the model only reads a few small rows of it per word.
 
-The full story is in the [paper](docs/paper/Strata-Paper.pdf) and the [details](docs/DETAILS.md).
+<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
+
+- **Guess, then check.** A small, fast helper built into the model guesses the next few words, and the big model
+  checks all the guesses in one go. It keeps the ones it agrees with and writes the next word itself - so one step
+  often produces several words. The helper only guesses - the big model decides every word - so you get the same
+  quality answer, 1.6-1.8x sooner.
+- **Long texts are read in big pieces** (up to 8,192 tokens - pieces of words - at a time), which is why a long
+  document or code base is read at over 1,000 tokens per second.
+
+Want the full picture? The [details](docs/DETAILS.md#how-it-works) explain every part and its numbers, and the
+[paper](docs/paper/Strata-Paper.pdf) tells the whole story, with the measurements behind it.
 
 ## Credits
 
@@ -138,3 +204,10 @@ The full story is in the [paper](docs/paper/Strata-Paper.pdf) and the [details](
 - Built with parts of [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) (MIT). Ideas from
   [Splash](https://github.com/incoai/splash), [ninfer](https://github.com/Neroued/ninfer) and
   [HyperQwen](https://github.com/syv-ai/HyperQwen). More in the [details](docs/DETAILS.md#credits-and-licenses).
+
+## License
+
+Strata is open source under the [MIT License](LICENSE). A few parts carry their own licenses: `third_party/ggml`
+(MIT, llama.cpp / ggml), the web app's font (SIL Open Font License 1.1) and the experimental speed projection's
+vector in `data/experimental-speed-projection` (Qwen Community License 1.0, from the model's activations). The
+models are not part of this repository; each model's own license applies to its files.

SETUP.bat

@@ -0,0 +1,3 @@
+@echo off
+rem Change settings or install another model (the same as START-HERE.bat --setup).
+call "%~dp0START-HERE.bat" --setup %*

bench/results/2026-09-27-cache-parity/README.md

@@ -0,0 +1,29 @@
+# 2026-09-27: expert cache on vs off (issue #23)
+
+Question: the engine used to warn at start that "the GPU hit path is NOT CORRECT ... any OUTPUT from it is not". Does
+the expert cache make the output worse?
+
+Method: teacher-forced. `strata --dump-logits` runs a text one token at a time and writes the logits at every
+position, so both runs see the same text and each position's next-token prediction can be compared:
+- cache off: no `--expert-cache`, every routed expert computed by the CPU pool;
+- cache on: `--expert-profile pack/profile-decode-8k.bin --expert-cache auto`, 5,130 of the most-used experts on the GPU.
+
+Q2_0, RTX 5070, the first 512 or 1,024 tokens of three frozen prompts.
+
+| Text | Same top-1 | Median KL | Perplexity off | Perplexity on | On - off (nats per token) |
+|---|---:|---:|---:|---:|---:|
+| Code agent (512) | 97.7% | 3.0e-04 | 2.672 | 2.630 | -0.016 +- 0.019 |
+| Long document (1,024) | 96.2% | 4.5e-03 | 29.844 | 29.843 | -0.000 +- 0.005 |
+| Chat (1,024) | 95.0% | 4.4e-03 | 18.380 | 18.301 | -0.004 +- 0.006 |
+| **All 2,557 tokens** | | | | | **-0.005 +- 0.005** |
+
+Reading:
+- The cache changes which of two nearly equal candidates wins at 2-5% of positions. Most of the flips have a top-2
+  margin under 0.5 logits (`logits-compare.txt`).
+- Neither side is more accurate: the perplexity of the actual text is the same, within 1 standard error.
+- A GPU expert and the CPU compute the same quantized expert with a different floating-point order. Both are an
+  equally good approximation of the model.
+
+The start-up warning dates from round 328, when the GPU hit path had a real fault (divergence from token 0). That fault
+was fixed long ago and is covered by `expert_parity`, `native_expert_parity` and the grouped kernels' tests. Since
+engine 0.1.8 the warning is replaced by one accurate line.

bench/results/2026-09-27-esp/README.md

@@ -0,0 +1,36 @@
+# 2026-09-27: the experimental-speed-projection control vector
+
+The package's `Qwen3.8-Flash-Next-experimental-speed-projection.gguf` (47 unit directions of 2,560, one per layer),
+applied as `--control-vector-scaled FILE:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir
+per-layer`: after layers 4-44, every hyper-connection stream `h -= (h . v) v`. The package describes it as a
+refusal-direction projection. Q2_0, RTX 5070, fixed experts (`--expert-cache 4000 --pcie-frac 0 --adapt-swaps 0`).
+
+## Kernel (`cvec_parity`)
+
+Project and add against a double-precision reference (max error 5e-7); with the switch off, the kernel's pending
+residual write is bitwise the fused hyper-connection read's own, so a loaded-but-off vector is the stock engine.
+
+## Teacher-forced, stock vs projection (`esp_kl.py`, per-token path)
+
+| Text | Tokens | Same top-1 | Mean KL | p99 KL | Perplexity stock | With the vector |
+|---|---:|---:|---:|---:|---:|---:|
+| Code | 511 | 92.8% | 0.098 | 1.69 | 2.594 | 2.982 |
+| Document | 1,023 | 89.8% | 0.058 | 0.84 | 30.299 | 30.984 |
+| Chat | 1,023 | 89.6% | 0.050 | 0.33 | 18.409 | 18.485 |
+| **All** | 2,557 | | **0.063** (max 4.10) | 1.00 | | |
+
+The package reports a mean KL of 0.186 on its own neutral text with UD-Q2_K_XL in llama.cpp.
+
+Cost on the same tokens: 44.34 -> 44.26, 46.74 -> 46.56, 47.02 -> 46.83 tok/s (0.2-0.4%).
+
+## Serve (`esp_serve.py`)
+
+- Loaded and switched off (`cvec=0`) vs an engine without it: 65 / 65 first tokens after prefixes (verify windows
+  and the batched prompt path) and three 256-token generations **token-identical**.
+- On, off, on again: identical (the switch drops the conversation cache, so nothing computed one way is reused the
+  other way).
+- The prompt paths steer like the per-token path: served-with-vector first tokens agree with the per-token path's
+  teacher-forced argmax with the vector 62 / 65, and with the stock one 57 / 65 (stock served: 64 / 65 and 57 / 65).
+- Decode tok/s of 3 x 256-token generations: stock 66.4, on 76.7 / 81.2. **Not a like-for-like comparison:** the
+  model writes different text with it on (on the code prompt it copied a long stretch of the source, which prompt
+  lookup drafts at 91% acceptance, where stock stopped after 78 tokens). On the same tokens it is 0.2-0.4% slower.

bench/results/2026-09-27-esp/esp_kl.py

@@ -0,0 +1,85 @@
+"""Teacher-forced logits, stock vs the experimental-speed-projection vector, on the same texts and fixed experts.
+
+    python esp_kl.py run      # the dumps (per-token path)
+    python esp_kl.py compare  # same top-1, mean/median KL(stock || projection), perplexities
+"""
+import json
+import os
+import subprocess
+import sys
+
+import numpy as np
+
+S = os.path.dirname(os.path.abspath(__file__))
+E = r"C:\Users\AI-Server\Desktop\Strata\Public\Engine"
+EXE = E + r"\build-dev\strata.exe"
+VEC = r"C:\Users\AI-Server\Desktop\Strata\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf"
+TEXTS = {"code": S + r"\p512.ids", "doc": S + r"\4k-02-long-doc-1k.ids", "chat": S + r"\4k-04-chat-1k.ids"}
+OUT = S + r"\esp-kl"
+ARMS = {"stock": [],
+        "proj": ["--control-vector-scaled", VEC + ":1.0", "--control-vector-layer-range", "4", "44",
+                 "--cvec-mode", "project", "--cvec-dir", "per-layer"]}
+
+
+def base_args():
+    c = json.load(open(S + r"\cfg-rel.json"))
+    a = list(c["args"])
+    i = a.index("--expert-cache")
+    a[i + 1] = "4000"
+    j = a.index("--prefill")          # the prompt one token at a time: a logits row per position
+    del a[j:j + 2]
+    return a + ["--pcie-frac", "0", "--adapt-swaps", "0"], c
+
+
+def run():
+    os.makedirs(OUT, exist_ok=True)
+    args, c = base_args()
+    env = dict(os.environ)
+    env["PATH"] = os.pathsep.join(c["lib_dirs"]) + os.pathsep + env["PATH"]
+    for name, path in TEXTS.items():
+        for arm, extra in ARMS.items():
+            out = f"{OUT}\\{name}-{arm}.bin"
+            if os.path.exists(out):
+                continue
+            cmd = [EXE, *args, *extra, "--tokens-file", path, "--max-new", "1", "--dump-logits", out]
+            with open(f"{OUT}\\{name}-{arm}.log", "w") as log:
+                r = subprocess.run(cmd, cwd=c["cwd"], stdout=log, stderr=subprocess.STDOUT, env=env)
+            print(name, arm, "exit", r.returncode, flush=True)
+
+
+def load(p):
+    h = np.fromfile(p, dtype=np.int32, count=2)
+    return np.memmap(p, dtype=np.float32, mode="r", offset=8, shape=(int(h[1]), int(h[0])))
+
+
+def lsm(x):
+    x = x.astype(np.float64)
+    m = x.max(axis=-1, keepdims=True)
+    return x - (m + np.log(np.exp(x - m).sum(axis=-1, keepdims=True)))
+
+
+def compare():
+    print(f"{'text':<6}{'tokens':>7}{'same top-1':>11}{'mean KL':>9}{'median KL':>11}{'p99 KL':>8}{'ppl stock':>10}{'ppl proj':>9}")
+    allkl = []
+    for name, path in TEXTS.items():
+        ids = [int(x) for x in open(path).read().replace(",", " ").split()]
+        a, b = load(f"{OUT}\\{name}-stock.bin"), load(f"{OUT}\\{name}-proj.bin")
+        n = min(a.shape[0], b.shape[0], len(ids) - 1)
+        kl, same, la, lb = [], 0, [], []
+        for r in range(0, n, 64):
+            pa, pb = lsm(a[r:min(n, r + 64)]), lsm(b[r:min(n, r + 64)])
+            kl += list((np.exp(pa) * (pa - pb)).sum(axis=1))
+            same += int((pa.argmax(1) == pb.argmax(1)).sum())
+            nxt = np.array(ids[r + 1:min(n, r + 64) + 1])
+            la += list(pa[np.arange(len(nxt)), nxt])
+            lb += list(pb[np.arange(len(nxt)), nxt])
+        kl = np.array(kl)
+        allkl += list(kl)
+        print(f"{name:<6}{n:>7}{100 * same / n:>10.1f}%{kl.mean():>9.3f}{np.median(kl):>11.4f}{np.percentile(kl, 99):>8.2f}"
+              f"{np.exp(-np.mean(la)):>10.3f}{np.exp(-np.mean(lb)):>9.3f}")
+    allkl = np.array(allkl)
+    print(f"all {len(allkl)} tokens: mean KL {allkl.mean():.3f}, max {allkl.max():.2f}, p99 {np.percentile(allkl, 99):.2f}")
+
+
+if __name__ == "__main__":
+    {"run": run, "compare": compare}[sys.argv[1]]()

bench/results/2026-09-27-esp/esp_serve.py

@@ -0,0 +1,129 @@
+"""Serve-mode checks of the experimental-speed-projection vector (exact mode: fixed experts, no PCIe share).
+
+    python esp_serve.py run ARM      # ARM = stock | vec; writes esp-serve/<arm>.json
+    python esp_serve.py compare
+
+Per arm: the first greedy token after many prefixes of three texts (short ones go through the verify windows, long
+ones through the batched prompt path; the conversation cache reuses the earlier prefix), and three 256-token greedy
+generations.  The vec arm runs everything with cvec=1, then cvec=0, then cvec=1 again.
+"""
+import json
+import os
+import sys
+import threading
+
+S = os.path.dirname(os.path.abspath(__file__))
+E = r"C:\Users\AI-Server\Desktop\Strata\Public\Engine"
+sys.path.insert(0, E)
+from serve.server import StrataEngine  # noqa: E402
+
+EXE = E + r"\build-dev\strata.exe"
+VEC = r"C:\Users\AI-Server\Desktop\Strata\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf"
+TEXTS = {"code": S + r"\p512.ids", "doc": S + r"\4k-02-long-doc-1k.ids", "chat": S + r"\4k-04-chat-1k.ids"}
+OUT = S + r"\esp-serve"
+VEC_ARGS = ["--control-vector-scaled", VEC + ":1.0", "--control-vector-layer-range", "4", "44",
+            "--cvec-mode", "project", "--cvec-dir", "per-layer"]
+
+
+def ids_of(name):
+    return [int(x) for x in open(TEXTS[name]).read().replace(",", " ").split()]
+
+
+def engine(extra):
+    c = json.load(open(S + r"\cfg-rel.json"))
+    a = list(c["args"])
+    a[a.index("--expert-cache") + 1] = "4000"
+    env = dict(os.environ)
+    env["PATH"] = os.pathsep.join(c["lib_dirs"]) + os.pathsep + env["PATH"]
+    return StrataEngine(EXE, a + ["--pcie-frac", "0", "--adapt-swaps", "0"] + extra, cwd=c["cwd"],
+                        log=OUT + r"\engine.log", env=env)
+
+
+def gen(eng, ids, n, sampling):
+    toks = [t for t in eng.generate(ids, n, sampling, threading.Event()) if t is not None]
+    return toks, dict(eng.last)
+
+
+def suite(eng, sampling):
+    out = {"first": {}, "long": {}}
+    for name in TEXTS:
+        ids = ids_of(name)
+        # short prefixes growing by 4 (each continues the live session through the verify windows), then long ones
+        # SHRINKING (no cached prefix, so each is read from 0 by the batched prompt path)
+        lens = list(range(32, 72, 4)) + list(range(min(len(ids) - 1, 700), 95, -48))
+        out["first"][name] = {L: gen(eng, ids[:L], 1, sampling)[0][0] for L in lens}
+    for name in TEXTS:
+        toks, last = gen(eng, ids_of(name)[:300], 256, sampling)
+        out["long"][name] = {"tokens": toks, "last": last}
+    return out
+
+
+def run(arm):
+    os.makedirs(OUT, exist_ok=True)
+    eng = engine(VEC_ARGS if arm == "vec" else [])
+    res = {"info": eng.info}
+    if arm == "stock":
+        res["stock"] = suite(eng, {})
+    else:
+        res["on"] = suite(eng, {"experimental_speed_projection": True})
+        res["off"] = suite(eng, {"experimental_speed_projection": False})
+        res["on2"] = suite(eng, {"experimental_speed_projection": True})
+    eng.close()
+    json.dump(res, open(f"{OUT}\\{arm}.json", "w"))
+    print(arm, "done", eng.info, flush=True)
+
+
+def compare():
+    import numpy as np
+    sys.path.insert(0, S)
+    from esp_kl import load, OUT as KL
+    st = json.load(open(OUT + r"\stock.json"))
+    ve = json.load(open(OUT + r"\vec.json"))
+    print("engine INFO (vec):", ve["info"])
+
+    def eq_first(a, b):
+        tot = same = 0
+        for name in a:
+            for L, t in a[name].items():
+                tot += 1
+                same += int(b[name][L] == t)
+        return same, tot
+
+    def eq_long(a, b):
+        return all(a[n]["tokens"] == b[n]["tokens"] for n in a), \
+            [(n, next((i for i, (x, y) in enumerate(zip(a[n]["tokens"], b[n]["tokens"])) if x != y), None)) for n in a]
+
+    print("1. vector loaded, switched off == no vector")
+    print("   first tokens: %d / %d identical" % eq_first(st["stock"]["first"], ve["off"]["first"]))
+    print("   256-token generations identical:", *eq_long(st["stock"]["long"], ve["off"]["long"]))
+    print("2. repeatable: on, off, on again")
+    print("   first tokens: %d / %d identical" % eq_first(ve["on"]["first"], ve["on2"]["first"]))
+    print("   256-token generations identical:", *eq_long(ve["on"]["long"], ve["on2"]["long"]))
+    print("3. the prompt paths against the per-token path's teacher-forced argmax (row L-1 predicts token L)")
+    for arm, first in (("stock", st["stock"]["first"]), ("on", ve["on"]["first"])):
+        for ref in ("stock", "proj"):
+            agree = tot = 0
+            for name in first:
+                d = load(f"{KL}\\{name}-{ref}.bin")
+                for L, t in first[name].items():
+                    L = int(L)
+                    if L - 1 < d.shape[0]:
+                        tot += 1
+                        agree += int(int(np.argmax(d[L - 1])) == t)
+            print(f"   served {arm:<5} vs per-token {ref:<5}: {agree} / {tot}")
+    print("4. how different the text is: 256-token generations, on vs stock, first difference")
+    print("  ", eq_long(st["stock"]["long"], ve["on"]["long"])[1])
+    print("5. decode speed (exact mode: fixed experts, no PCIe share - not the normal setup)")
+    for label, r in (("stock", st["stock"]), ("vec off", ve["off"]), ("vec on", ve["on"]), ("vec on 2", ve["on2"])):
+        g = sum(v["last"]["generated"] for v in r["long"].values())
+        ms = sum(v["last"]["decode_ms"] for v in r["long"].values())
+        acc = sum(v["last"].get("drafts_accepted", 0) for v in r["long"].values())
+        off = sum(v["last"].get("drafts_offered", 0) for v in r["long"].values())
+        print(f"   {label:<9} {g / ms * 1000:6.1f} tok/s  drafts accepted {acc}/{off}")
+
+
+if __name__ == "__main__":
+    if sys.argv[1] == "run":
+        run(sys.argv[2])
+    else:
+        compare()

bench/results/2026-09-27-esp/serve-checks.txt

@@ -0,0 +1,20 @@
+1. vector loaded, switched off == no vector
+   first tokens: 65 / 65 identical
+   256-token generations identical: True [('code', None), ('doc', None), ('chat', None)]
+2. repeatable: on, off, on again
+   first tokens: 65 / 65 identical
+   256-token generations identical: True [('code', None), ('doc', None), ('chat', None)]
+3. the prompt paths against the per-token path's teacher-forced argmax (row L-1 predicts token L)
+   served stock vs per-token stock: 64 / 65
+   served stock vs per-token proj : 57 / 65
+   served on    vs per-token stock: 57 / 65
+   served on    vs per-token proj : 62 / 65
+4. how different the text is: 256-token generations, on vs stock, first difference
+   [('code', 11), ('doc', 6), ('chat', 3)]
+5. decode speed (exact mode: fixed experts, no PCIe share - not the normal setup)
+   stock       66.4 tok/s  drafts accepted 256/309
+   vec off     65.7 tok/s  drafts accepted 251/318
+   vec on      76.7 tok/s  drafts accepted 366/425
+   vec on 2    81.2 tok/s  drafts accepted 366/423
+
+[exited with code 0]

bench/results/2026-09-27-esp/teacher-forced.txt

@@ -0,0 +1,6 @@
+text   tokens same top-1  mean KL  median KL  p99 KL ppl stock ppl proj
+code      511      92.8%    0.098     0.0031    1.69     2.594    2.982
+doc      1023      89.8%    0.058     0.0271    0.84    30.299   30.984
+chat     1023      89.6%    0.050     0.0265    0.33    18.409   18.485
+all 2557 tokens: mean KL 0.063, max 4.10, p99 1.00
+

bench/results/2026-09-27-kv-q4/README.md

@@ -0,0 +1,60 @@
+# 2026-09-27: Q4_0 KV with Hadamard rotation (PR #21, `--kv q4_0`)
+
+PR #21 (code-martin) adds a 4-bit K/V format: an orthonormal 256-point Walsh-Hadamard rotation of K, V and Q, then
+ggml `q4_0` blocks (576 B per cell and layer, vs 1,056 B in int8).
+
+## Integration fixes made on merge
+
+- **Q4 + KV streaming (`--kv-resident`).** The PR had no host copy for Q4 and skipped the staging, so with streaming
+  on, the prompt path wrote K/V through the residency map's -1 entries (out-of-bounds writes).
+  - Q4 is now a third format in `kv_stream`: host copy, block residency, prompt-path staging, MTP ring and STATE_HASH.
+  - `kv_stream_parity` checks it bitwise, streamed vs resident, with ring restore.
+- **Duplicate attention.** The per-token decode path ran attention twice, the first time before the selection's
+  blocks were made resident; it now runs once.
+- **Tie-break.** The group maximum's tie-break is now deterministic across lanes: equal |x| with opposite signs
+  could give one block two scales.
+- **Unrelated changes dropped:** duplicate path fixes (already in 0.1.7), personal `.gitignore` entries, and a
+  re-encoded CMakeLists.
+
+`kv_q4_parity` (the PR's test) and all 70 ctest tests pass.
+
+## Quality (teacher-forced, `tools/kv_precision_compare.py`)
+
+Same text, same fixed expert set (Q2_0, 3,000 slots, `--adapt-swaps 0`); only `--kv` differs. fp16 is the reference.
+
+| Text | KV | Same top-1 as fp16 | Mean KL | Perplexity | vs fp16 (nats/token) |
+|---|---|---:|---:|---:|---:|
+| code agent (512) | int8 | 98.4% | 0.014 | 2.649 | -0.011 +- 0.006 |
+| | q4_0 | 95.7% | 0.030 | 2.598 | -0.031 +- 0.014 |
+| document (1,024) | int8 | 95.6% | 0.009 | 30.18 | +0.012 +- 0.005 |
+| | q4_0 | 91.3% | 0.039 | 32.26 | **+0.078 +- 0.011** |
+| chat (1,024) | int8 | 95.5% | 0.009 | 18.24 | -0.003 +- 0.005 |
+| | q4_0 | 90.7% | 0.028 | 18.52 | +0.012 +- 0.008 |
+| document, 8K (every 8th position) | int8 | 92.3% | 0.114 | 7.63 | -0.013 +- 0.020 |
+| | q4_0 | 88.0% | 0.297 | 8.65 | **+0.112 +- 0.037** |
+
+- **int8 is indistinguishable from fp16** (within about 2 standard errors everywhere).
+- **q4_0 costs measurable precision, and more as the context grows:**
+  - document perplexity +8% at 1K and +12% at 8K;
+  - the KL from fp16 is 2.6-4.2x int8's;
+  - code and chat are near-neutral.
+- **Needles with `--kv q4_0 --kv-resident 32768`: 5/5** (1K, 32K, 128K at 10% and 90% depth, 262K;
+  `needles-q4.*`). Retrieval holds.
+
+## Speed and memory (Q2_0, 128K, matrix bench, `--kv-resident 32768`)
+
+| KV | Expert slots | Decode tok/s | Tokens/round | Rounds/s | K/V in RAM |
+|---|---:|---:|---:|---:|---:|
+| int8 (`kv-stream/matrix-stream`) | 4,044 | 62.4 | 2.52 | 24.8 | 1.52 GiB |
+| q4_0 (`matrix-q4`) | 4,190 | 67.4 | 2.62 | 25.7 (+3.7%) | 0.83 GiB |
+
+Without streaming (contexts up to 64K, or too little RAM for it) q4_0 halves the whole KV cache in VRAM instead.
+
+## Decision
+
+`--kv q4_0` is a user option:
+- `START-HERE.bat --kv q4_0`, or the question setup asks above 8K context;
+- or `--kv q4_0` in a `strata-*.json`.
+
+int8 stays the default: every published quality number was measured with it, and q4_0's precision cost is real,
+though it passes the needles.

bench/results/2026-09-28-coder/README.md

@@ -0,0 +1,44 @@
+# The GSQ-RCO Coder (256 of 512 experts) on Strata
+
+RTX 5070 12 GB, Ryzen 5 7600, 64 GB DDR5-5200, Windows 10. Engine 0.1.15 + PR #54 (the expert count from the model
+file), the settings setup writes (`--prefill auto`, 8-bit KV above 4K, KV streaming from 64K, MTP, greedy, 256 tokens),
+`data/expert-profile-coder.bin`, the same code-agent prompts as `2026-09-28-speed-0114`.
+
+| | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Prompt, tokens/s | 599 | 1,152 | 1,298 | 1,350 | 1,266 | 1,034 |
+| Output, tokens/s | 53.3 | 50.6 | 53.3 | 50.8 | 44.0 | 42.8 |
+| Draft acceptance | 0.720 | 0.707 | 0.785 | 0.696 | 0.619 | 0.684 |
+| Experts in VRAM | 2,570 | 2,615 | 2,418 | 2,382 | 2,326 | 2,208 |
+
+Every run in [`matrix.json`](matrix.json). The 128K and 262K answers were read: coherent (they list and discuss the
+files of the prompt).
+
+## Why it is fast
+
+The kept experts are stored like IQ3_S's (IQ2_S / IQ3_XXS / IQ3_S / IQ4_XS gate and up, IQ4_NL / Q2_0 down; "IQ1_M" is
+the release's name for 1.89 bits per *original* parameter), so a token costs about what IQ3_S costs where the expert is
+on the CPU. But there are half as many experts: the 23 GB of them pin completely (`cudaHostRegister` of the whole
+arena), a 12 GB card holds ~20% of them instead of ~10%, and the prompt path streams half as many.
+
+## The expert profile
+
+`data/expert-profile-coder.bin` (48 x 256, all 12,288 pairs ranked): the shipped 48 x 512 ranking, each pair mapped to
+its kept expert's new index through section 2 of the release's `tensor-allocation/*.rco-allocation.txt` (the kept
+experts in their original order; gate, up and down agree on which were kept), pruned pairs dropped. 5,492 of the
+shipped top 8,000 survive (as PR #54 found). With 2,537 slots on the 12 GB card, 72% of the expert reads of a 4K code
+prompt hit the GPU - a wrong index mapping would give about the slot share, 21%.
+
+## Images
+
+The Coder ships the original's vision encoder (the same file). With `--vision` on the GPU it described a Strata
+Monitor screenshot correctly, windows, labels and the numbers of the dashboard and the requests table (1,079 prompt
+tokens with the image, 2,012 generated at ~50 tokens/s).
+
+## The experimental speed projection
+
+The vector was made from the full model's activations; the Coder has the same residual stream. On the Coder it loads
+(41 layers steered), costs no measurable speed (47.7-55.7 tokens/s with it on or off), and both settings write correct
+answers to three coding prompts. What it is for - fewer refusals - was not measured. (Two runs of the same prompt with
+it off also differ after ~130-200 characters in serve mode, so this check cannot show how far it moves ordinary text;
+the teacher-forced method of `2026-09-27-esp` needs the token-by-token prompt path, which native packs do not run.)

bench/results/2026-09-28-coder/matrix.json

@@ -0,0 +1,86 @@
+[
+ {
+  "model": "CODER",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 18.0,
+  "prefill_tok_s": 599.05,
+  "ttft_s": 2.54,
+  "decode_tok_s": 53.33,
+  "decoded": 256,
+  "spec_accept": 0.72,
+  "tokens_per_round": 2.86,
+  "vram_slots": 2570
+ },
+ {
+  "model": "CODER",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 19.6,
+  "prefill_tok_s": 1152.17,
+  "ttft_s": 3.94,
+  "decode_tok_s": 50.6,
+  "decoded": 256,
+  "spec_accept": 0.707,
+  "tokens_per_round": 2.88,
+  "vram_slots": 2615
+ },
+ {
+  "model": "CODER",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 40.7,
+  "prefill_tok_s": 1297.8,
+  "ttft_s": 25.17,
+  "decode_tok_s": 53.34,
+  "decoded": 256,
+  "spec_accept": 0.785,
+  "tokens_per_round": 3.02,
+  "vram_slots": 2418
+ },
+ {
+  "model": "CODER",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 75.5,
+  "prefill_tok_s": 1350.09,
+  "ttft_s": 49.01,
+  "decode_tok_s": 50.84,
+  "decoded": 256,
+  "spec_accept": 0.696,
+  "tokens_per_round": 2.61,
+  "vram_slots": 2382
+ },
+ {
+  "model": "CODER",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 121.5,
+  "prefill_tok_s": 1266.31,
+  "ttft_s": 102.95,
+  "decode_tok_s": 44.03,
+  "decoded": 256,
+  "spec_accept": 0.619,
+  "tokens_per_round": 2.33,
+  "vram_slots": 2326
+ },
+ {
+  "model": "CODER",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 280.7,
+  "prefill_tok_s": 1033.91,
+  "ttft_s": 252.92,
+  "decode_tok_s": 42.8,
+  "decoded": 256,
+  "spec_accept": 0.684,
+  "tokens_per_round": 2.6,
+  "vram_slots": 2208
+ }
+]

bench/results/2026-09-28-prefill-speed/README.md

@@ -0,0 +1,47 @@
+# Faster prompt reading (engine 0.1.13)
+
+RTX 5070 12 GB, Ryzen 5 7600, 64 GB DDR5. 32K-token prompt (`bench/prompts/long/32k-01-code-agent.ids`), one-shot
+engine, `--expert-cache auto`, greedy. Numbers are prompt tokens/s.
+
+## What changed, step by step
+
+| Step | What | Result |
+|---|---|---|
+| Buffers | The attention half and the MoE half of a layer share one scratch region (~260 KB less per prompt token) | bit-identical; chunk 4096 borrows 1.78 GiB instead of 2.84 |
+| PLE rows | The next chunk's PLE rows are read from the model file on a thread while the GPU works | bit-identical; 762 -> 790 at chunk 4096 |
+| Chunk size | `--prefill auto`: the largest of 8192 / 6144 / 4096 / ... whose buffers fit in the cache slots it may borrow; a request borrows only what its own prompt needs | Q2_0: 4096 791, 6144 878, 8192 973 |
+| Host copies | Experts the arena could not pin are copied to pinned buffers by helper threads instead of the launching thread | bit-identical; IQ3_S at 4096: 551 -> 652 |
+| PLE block | Computed for the whole chunk (two GEMMs) instead of token by token | Q2_0 981 -> 1053, IQ3_S 831 -> 940 |
+| Experts | llama.cpp's MMQ kernels (weights stay quantized, int8 tensor cores) instead of dequantize-to-FP16 + cuBLAS | Q2_0 1052 -> 1130, IQ3_S (8192) 932 -> 1014 |
+| Streaming | At chunks of 2048+ every non-resident expert streams in a fixed order through a ring, so the next layer's arrive while the current layer's attention runs | bit-identical; Q2_0 1130 -> 1290, IQ3_S -> 1208 |
+
+Tried and dropped: cuBLAS grouped GEMMs for the experts (faster in isolation, slower inside the engine: 770 vs 790).
+
+## Before and after
+
+| | engine 0.1.12 (`--prefill 2048`) | engine 0.1.13 (`--prefill auto`) |
+|---|---:|---:|
+| Q2_0, 32K prompt | 572 | 1,290 |
+| IQ3_S, 32K prompt | 383 | 1,208 |
+| Server, Q2_0, 128K context, 999-token prompt | 353 | 438 |
+| Server, 6,927-token prompt | 529 | 1,077 |
+| Server, 28,584-token prompt | 584 | 1,249 |
+
+Output speed is unchanged (the decode path is the same). Short prompts (a few hundred tokens) go through the verify
+windows as before.
+
+## Quality
+
+- "Bit-identical" steps were checked with the GDN state hash after the 32K prompt (all 36 recurrent layers).
+- The steps that change rounding (the batched PLE block, MMQ's 8-bit activations) were compared teacher-forced: the
+  32K prompt is read by the prompt path up to its last 256 tokens, those go through the token path with their logits
+  dumped. At 32K every rounding change moves the predictions about this much - even a different chunk size of the
+  unchanged engine - so that is the yardstick:
+
+| Against the 8192-chunk FP16 path | Same top-1 | Mean KL | Perplexity (reference 9.21) |
+|---|---:|---:|---:|
+| Chunk 6144, same code (the yardstick) | 89.8% | 0.33 | 9.18 |
+| MMQ experts | 85.9% | 0.38 | 8.63 |
+
+- Needles (the answer hidden at a given depth): 5 of 5 at 1K, 32K, 128K (two depths) and 262K, Q2_0, KV streaming.
+- A rerun of the same configuration is bit-identical.

bench/results/2026-09-28-speed-0114/README.md

@@ -0,0 +1,32 @@
+# Speed with engine 0.1.14
+
+RTX 5070 12 GB (PCIe 5.0 x16), Ryzen 5 7600, 64 GB DDR5-5200, Windows 10. The ready-made 0.1.14 engine, one-shot runs
+with what setup writes: `--prefill auto`, the shipped expert profile, `--expert-cache auto`, 8-bit KV above 4K, KV
+streaming (`--kv-resident 32768`) from 64K, MTP (`--spec 4 --spec-min-p 0.5`), greedy, 256 generated tokens. One
+code-agent prompt per length (`bench/prompts/*/…-code-agent.ids`, the same prompts as `2026-09-24-final`).
+
+The tables: [`matrix.md`](matrix.md) (prompt and output speed, draft acceptance, experts in VRAM); every run in
+[`matrix.json`](matrix.json).
+
+- **IQ2_XS** was measured with Swift 1.5's IQ2_XS (the original's file was not on the machine; the two run at the same
+  speed: 4K 465 vs 467 prompt tokens/s in `2026-09-24`).
+- **IQ3_S at 128K** comes from a first pass during which a large download was running (the clean rerun of that row
+  was stopped when Windows ran short of memory); the other rows are from runs on an otherwise idle machine.
+- **IQ3_XXS and IQ3_S at 262K** are not measured: with 43 / 50 GB of experts, a 260K context exceeds 64 GB of RAM.
+
+## Against 0.1.12
+
+Prompts are 1.2x (1K) to 2.4x (32K-128K) faster: the 0.1.13 prompt path (chunks of up to 8,192 tokens, MMQ experts,
+streamed expert copies; see `2026-09-28-prefill-speed`).
+
+Output speed moves with the text: speculative decoding is faster when more drafts are accepted, and the new prompt
+path's rounding changes the answer, so the same prompt can land a few percent either way (Q2_0 32K: acceptance 0.865
+then, 0.709 now). The engine itself is not slower: on the 4K prompt, run back to back on the same machine state,
+0.1.14 writes 88.5 tokens/s and 0.1.12 85.7 (3 runs each, same free VRAM).
+
+## Found on the way
+
+IQ3_XXS at 64K and 128K ended with `out of memory: cudaFuncSetAttribute` in the MMQ launch: CUDA loads a kernel's code
+when it is first used, and mid-prompt there was no VRAM left for it. Loading every kernel at start
+(`CUDA_MODULE_LOADING=EAGER`, which the engine now sets itself) fixes it for ~30 MB of VRAM; the IQ3_XXS 64K/128K rows
+were measured with that fix.

bench/results/2026-09-28-speed-0114/matrix.json

@@ -0,0 +1,322 @@
+[
+ {
+  "model": "Q2_0",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 41.0,
+  "prefill_tok_s": 494.43,
+  "ttft_s": 3.55,
+  "decode_tok_s": 84.26,
+  "decoded": 256,
+  "spec_accept": 0.709,
+  "tokens_per_round": 2.72,
+  "vram_slots": 4176
+ },
+ {
+  "model": "Q2_0",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 41.2,
+  "prefill_tok_s": 1006.63,
+  "ttft_s": 5.05,
+  "decode_tok_s": 90.26,
+  "decoded": 256,
+  "spec_accept": 0.846,
+  "tokens_per_round": 3.23,
+  "vram_slots": 4123
+ },
+ {
+  "model": "Q2_0",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 63.1,
+  "prefill_tok_s": 1307.6,
+  "ttft_s": 25.63,
+  "decode_tok_s": 73.61,
+  "decoded": 256,
+  "spec_accept": 0.709,
+  "tokens_per_round": 2.68,
+  "vram_slots": 3833
+ },
+ {
+  "model": "Q2_0",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 89.0,
+  "prefill_tok_s": 1293.5,
+  "ttft_s": 51.12,
+  "decode_tok_s": 69.02,
+  "decoded": 256,
+  "spec_accept": 0.679,
+  "tokens_per_round": 2.62,
+  "vram_slots": 3780
+ },
+ {
+  "model": "Q2_0",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 147.4,
+  "prefill_tok_s": 1208.46,
+  "ttft_s": 107.81,
+  "decode_tok_s": 67.21,
+  "decoded": 256,
+  "spec_accept": 0.691,
+  "tokens_per_round": 2.64,
+  "vram_slots": 3697
+ },
+ {
+  "model": "Q2_0",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 309.8,
+  "prefill_tok_s": 966.84,
+  "ttft_s": 270.35,
+  "decode_tok_s": 60.3,
+  "decoded": 256,
+  "spec_accept": 0.623,
+  "tokens_per_round": 2.43,
+  "vram_slots": 3525
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 40.4,
+  "prefill_tok_s": 461.28,
+  "ttft_s": 2.28,
+  "decode_tok_s": 74.38,
+  "decoded": 256,
+  "spec_accept": 0.772,
+  "tokens_per_round": 2.84,
+  "vram_slots": 4386,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 41.8,
+  "prefill_tok_s": 810.71,
+  "ttft_s": 4.48,
+  "decode_tok_s": 73.84,
+  "decoded": 256,
+  "spec_accept": 0.79,
+  "tokens_per_round": 3.24,
+  "vram_slots": 4336,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 65.0,
+  "prefill_tok_s": 1237.98,
+  "ttft_s": 25.59,
+  "decode_tok_s": 71.49,
+  "decoded": 256,
+  "spec_accept": 0.856,
+  "tokens_per_round": 3.28,
+  "vram_slots": 4057,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 97.1,
+  "prefill_tok_s": 1135.63,
+  "ttft_s": 56.59,
+  "decode_tok_s": 64.29,
+  "decoded": 256,
+  "spec_accept": 0.707,
+  "tokens_per_round": 2.76,
+  "vram_slots": 4005,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 158.0,
+  "prefill_tok_s": 1070.61,
+  "ttft_s": 120.1,
+  "decode_tok_s": 59.76,
+  "decoded": 256,
+  "spec_accept": 0.726,
+  "tokens_per_round": 2.69,
+  "vram_slots": 3923,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 332.4,
+  "prefill_tok_s": 886.2,
+  "ttft_s": 293.43,
+  "decode_tok_s": 52.83,
+  "decoded": 256,
+  "spec_accept": 0.628,
+  "tokens_per_round": 2.53,
+  "vram_slots": 3565,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 44.1,
+  "prefill_tok_s": 414.55,
+  "ttft_s": 2.54,
+  "decode_tok_s": 60.26,
+  "decoded": 256,
+  "spec_accept": 0.864,
+  "tokens_per_round": 3.52,
+  "vram_slots": 3304,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 45.8,
+  "prefill_tok_s": 769.57,
+  "ttft_s": 4.73,
+  "decode_tok_s": 62.07,
+  "decoded": 256,
+  "spec_accept": 0.855,
+  "tokens_per_round": 3.24,
+  "vram_slots": 3263,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 69.7,
+  "prefill_tok_s": 1107.75,
+  "ttft_s": 28.61,
+  "decode_tok_s": 51.4,
+  "decoded": 256,
+  "spec_accept": 0.678,
+  "tokens_per_round": 2.6,
+  "vram_slots": 3031,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 104.9,
+  "prefill_tok_s": 1065.31,
+  "ttft_s": 61.77,
+  "decode_tok_s": 49.95,
+  "decoded": 256,
+  "spec_accept": 0.741,
+  "tokens_per_round": 2.94,
+  "vram_slots": 2965,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 172.1,
+  "prefill_tok_s": 1015.11,
+  "ttft_s": 128.12,
+  "decode_tok_s": 45.78,
+  "decoded": 256,
+  "spec_accept": 0.665,
+  "tokens_per_round": 2.35,
+  "vram_slots": 2901,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 56.2,
+  "prefill_tok_s": 396.13,
+  "ttft_s": 4.14,
+  "decode_tok_s": 51.52,
+  "decoded": 256,
+  "spec_accept": 0.826,
+  "tokens_per_round": 3.12,
+  "vram_slots": 2617
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 57.5,
+  "prefill_tok_s": 736.9,
+  "ttft_s": 6.35,
+  "decode_tok_s": 51.6,
+  "decoded": 256,
+  "spec_accept": 0.747,
+  "tokens_per_round": 2.89,
+  "vram_slots": 2580
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 82.5,
+  "prefill_tok_s": 1069.83,
+  "ttft_s": 31.33,
+  "decode_tok_s": 48.19,
+  "decoded": 256,
+  "spec_accept": 0.772,
+  "tokens_per_round": 2.94,
+  "vram_slots": 2383
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 115.0,
+  "prefill_tok_s": 1069.8,
+  "ttft_s": 61.47,
+  "decode_tok_s": 48.75,
+  "decoded": 256,
+  "spec_accept": 0.786,
+  "tokens_per_round": 2.9,
+  "vram_slots": 2347
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 191.8,
+  "prefill_tok_s": 931.09,
+  "ttft_s": 138.11,
+  "decode_tok_s": 40.48,
+  "decoded": 256,
+  "spec_accept": 0.661,
+  "tokens_per_round": 2.71,
+  "vram_slots": 2310,
+  "note": "first pass (a download was running)"
+ }
+]

bench/results/2026-09-28-speed-0114/matrix.md

@@ -0,0 +1,35 @@
+### Prompt processing (tok/s)
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 494 | 1,007 | 1,308 | 1,294 | 1,208 | 967 |
+| IQ2_XS | 461 | 811 | 1,238 | 1,136 | 1,071 | 886 |
+| IQ3_XXS | 415 | 770 | 1,108 | 1,065 | 1,015 | - |
+| IQ3_S | 396 | 737 | 1,070 | 1,070 | 931 | - |
+
+### Output (tok/s)
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 84.3 | 90.3 | 73.6 | 69.0 | 67.2 | 60.3 |
+| IQ2_XS | 74.4 | 73.8 | 71.5 | 64.3 | 59.8 | 52.8 |
+| IQ3_XXS | 60.3 | 62.1 | 51.4 | 50.0 | 45.8 | - |
+| IQ3_S | 51.5 | 51.6 | 48.2 | 48.8 | 40.5 | - |
+
+### Draft acceptance
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 0.709 | 0.846 | 0.709 | 0.679 | 0.691 | 0.623 |
+| IQ2_XS | 0.772 | 0.790 | 0.856 | 0.707 | 0.726 | 0.628 |
+| IQ3_XXS | 0.864 | 0.855 | 0.678 | 0.741 | 0.665 | - |
+| IQ3_S | 0.826 | 0.747 | 0.772 | 0.786 | 0.661 | - |
+
+### Experts in VRAM
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 4,176 | 4,123 | 3,833 | 3,780 | 3,697 | 3,525 |
+| IQ2_XS | 4,386 | 4,336 | 4,057 | 4,005 | 3,923 | 3,565 |
+| IQ3_XXS | 3,304 | 3,263 | 3,031 | 2,965 | 2,901 | - |
+| IQ3_S | 2,617 | 2,580 | 2,383 | 2,347 | 2,310 | - |

bench/results/2026-09-29-layer-split/README.md

@@ -0,0 +1,84 @@
+# Layer split across GPUs (2026-09-29)
+
+**Rig (a community member's PC, "RIBPC"):**
+- Ryzen 9 9950X3D, 62 GB DDR5, Windows 11, CUDA 13.1.
+- RTX 5080 16 GB (x16), RTX 3090 24 GB (x4 gen4, ~6 GB/s), RTX 2080 Ti 22 GB (x4 gen3, Turing sm_75; the
+  experimental `STRATA_EXPERIMENTAL_SM75` build).
+
+**Model:** the Coder (GSQ-RCO IQ1_M, 48 layers x 256 experts = 12,288 expert pairs, ~23 GB of experts), 32K
+context, int8 KV, `--prefill auto`, `--spec 4`. It is the only model that fits the rig's RAM and disk.
+
+**Tools:** `rib_pp.py` (needle prompts at 2.7K/8K/16K/28K tokens with a code word at 50% depth, a 2-turn chat that
+must reuse its checkpoint, greedy decode of a story and a code prompt) and `rib_norepeat.py` (the Phase A
+no-repeat and sanity cases). Raw results are in `data/`.
+
+## Correctness
+
+| Check | Result |
+|---|---|
+| This branch vs the 0.1.20 release, one GPU (5080), fixed cache, all 10 cases | **10/10 byte-identical** |
+| This branch, one GPU vs the split on the same GPU (K=24, `--split-device 0`) | **10/10 byte-identical**: the hand-off is exact |
+| 5080 + 3090 (auto) and 5080 + 2080 Ti + 3090 (auto): no-repeat under presence 100 (greedy and sampled), 144 sanity | all pass, 0 repeats |
+| Needles at 2.7K-28K tokens, every configuration | all found |
+| 2-turn chat (the checkpoint restored on every GPU) | reused 6,631-6,632 of 6,660 tokens, answer right |
+
+Across GPUs the text is not byte-identical to one GPU: experts the CPU computed on one card run on a GPU in the split,
+and they round differently (the same effect as `--expert-cache`, bench/results/2026-09-27-cache-parity).
+
+## Speed
+
+**Prompt tok/s** at 16K / 28K tokens, and **decode tok/s** for story / code (greedy; three repeats averaged
+in the K sweep):
+
+| Configuration | Split | Prompt 16K | Prompt 28K | Decode story | Decode code | Decode hits |
+|---|---|---|---|---|---|---|
+| 0.1.20, 5080 alone | - | 2,045 | 2,005 | 83.2 | 88.2 | 70-84% |
+| this branch, 5080 alone | - | 2,017 | 1,974 | 83.1 | 88.0 | 70-84% |
+| this branch, 5080 alone, 3 repeats (adaptive swaps learn the prompt) | - | 1,726 | 1,970 | 86.7 | 105.2 | 91-92% |
+| 3090 alone | - | 1,253 | 1,211 | 71.7 | 92.9 | 86-93% |
+| 5080 + 3090 | K=18 | 1,547 | 1,694 | 77.2 | 103.8 | 99.9% |
+| 5080 + 3090 | K=22 | 1,840 | 2,074 | 80.2 | 106.4 | 99.8% |
+| 5080 + 3090 | K=24 | 1,978 | 2,258 | 83.5 | 109.5 | 99.5% |
+| **5080 + 3090** | **K=26** | **2,039** | **2,357** | **83.8** | **109.7** | 99.3% |
+| 5080 + 3090 | K=28 | 2,011 | 2,287 | 83.5 | 112.5 | 99.0% |
+| 5080 + 3090 | K=30 | 1,935 | 2,218 | 81.1 | 109.6 | 98% |
+| 5080 + 3090 | K=32 | 1,818 | 2,063 | 83.0 | 108.7 | 98% |
+| 5080 + 3090 | auto (K=22) | 2,037 | 2,073 | 79.5 | 108.9 | 99.7% |
+| 3090 + 5080 | auto (K=30) | 1,672 | 1,876 | 90.3 | 105.1 | 99.4% |
+| 5080 + 2080 Ti + 3090 | auto (22, 23) | 1,687 | 1,847 | 67.6 | 89.6 | 99.7% |
+| 5080 + 3090 + 2080 Ti | first auto (12, 33) | 1,148 | 1,144 | 51.9 | 69.7 | 100% |
+| 5080 + 2080 Ti | first auto | 914 | 906 | 45.5 | 58.0 | 97% |
+
+What decides it:
+
+- **The best two-GPU split for the Coder is K=26-28.** Prompts run +18-20% over the 5080 alone at 16-28K;
+  decode is on par for story and +4-7% for code.
+- **Residency saturates quickly: the per-layer GPU speed decides.** From K=18 up, both caches hold 98-100% of
+  the routed experts. Moving layers to the faster card is what helps: the 3090 costs 0.50 ms per layer per
+  window, the 5080 0.33 (`STRATA_SPLIT_TIMING`, the "wait for the GPU" per stage).
+- **Decode is bounded by that GPU time.** One 5080 with its cache adapted spends 16.3 ms waiting for its GPU and
+  5.5 ms in the CPU pool per window. The split removes the pool (0.8-1.0 ms) but adds the slower card's layers:
+  7.9 + 12.1 ms at K=24. The per-layer GPU half (kernels plus the host round trip that publishes each layer's
+  expert plan) is the next thing to cut, for one GPU as well (perf-review-plan E-6).
+- **A third, slower card costs more than it holds.** The 2080 Ti's one-layer stage took 3.1 ms per window.
+  For a model that already fits two cards, leave it out. It pays only when the two faster cards cannot hold most
+  of the routed experts (a bigger model; not testable on this rig's 62 GB of RAM).
+- **Prompt chunk:** 1024 instead of 2048 frees ~0.7 GB per card, but prompts read 7-12% slower.
+
+## Auto placement
+
+`--layer-split auto` predicts one decode window's time for every placement (every placement up to three GPUs) and
+keeps the fastest:
+
+- **per-layer time** from SMs x clock (0.33 ms for an RTX 5080);
+- **cache contents:** each card's layers' profiled pairs, hottest first, as the fill does;
+- **routed mass** of rank r taken as (r+1)^-1.2 (fits the sweep's hit rates within 0.1-1%);
+- **missed mass** costing 190 ms per unit (fitted to the sweep's best K; `STRATA_SPLIT_MISS_MS` overrides).
+
+It picked K=22 on the 5080 + 3090: within 0-12% of the best fixed split on prompts and on par on decode. It gives
+a slow third card a single layer.
+
+## Reproduce
+
+On a rig: the `strata-remote` tools, then `rib_matrix.bat`, `rib_ksweep.bat`, `rib_validate.bat` (in the dev
+notes, not in this repository).

bench/results/2026-09-29-layer-split/data/norepeat-nr-v-x2-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 193,
+  "tok_s": 74.6,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, gripped the iron railing as he ascended the final flight of steps to the lantern room.\n\nBelow him, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t flip a switch. He didn\u2019t press a button. He reached for the heavy brass lever that had worn smooth by his palm over the years. With a grunt that matched the strain of his aging muscles, he pulled it down.\n\n*Clunk.*\n\nThe"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 181,
+  "tok_s": 105.5,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 85.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 94.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 105,
+  "tok_s": 67.9,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 147,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 103,
+  "tok_s": 63.2,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to that date and time for context because you asked me about a lighthouse keeper but gave no specifics beyond \u201cshort story,\u201d so let\u2019s set it: isolated coastal station off the Oregon coast during winter storm season, 1987.\n\nMy name is Elias Vane (I made this fictional; real keepers often kept journals signed with initials only), and I\u2019m writing from memory because my logbook got waterlogged when a rogue wave breached lower storage room at dawn\u2014so here\u2019s what happened:<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 109,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 56,
+  "tok_s": 54.1,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they settled into silence with every storm break out of three hundred days here being storms that broke something else again but mostly kept me warm because I burned coal and wood whenever we had any left over from last winter when even then if you\u2019re lucky enough to get through it all without freezing yourself so hard your own thoughts become ice too\n\nBut Elias knew better than anyone: fear was useful only until dawn<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 92,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 48,
+  "tok_s": 64.7,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nHippopotamus          \nRhinoceros            \nCamel              \nCheetah             \nLeopard           \nPanther         \nJaguar                \nBison               \nElk                 \nDeer                  \nFox                   \nWolf                    \nBear                     \nLlama                       \nDonkey                        \nGoat                      \nSheep                         \nCow                           \nPig                          \nRabbit                            \nCat                              \nDog<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 14,
+  "tok_s": 71.7,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 53,
+  "drafts_offered": 134,
+  "tok_s": 66.1,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story we tell ourselves through our relationship** between landmasses separated by vast expansiveness that defines both isolation from other worlds (and thus their own unique evolutionary paths) as well than connection across them via currents wind patterns tides depth pressure temperature salinity biodiversity geology climate change economic trade routes ecological systems cultural exchange political boundaries scientific discovery technological advancement environmental degradation conservation efforts maritime archaeography historical cartographic evolution legal frameworks international law enforcement mechanisms insurance liability regimes regulatory compliance standards safety protocols navigation aids communication networks emergency response strategies search-and-rescue operations salvage rights ownership disputes heritage protection zones sustainable harvesting practices aquaculture development infrastructure engineering projects coastal management policies urban planning regulations zoning ordinances building codes architectural design guidelines aesthetic considerations functional requirements accessibility provisions energy efficiency measures thermal insulation techniques moisture control methods ventilation optimization lighting solutions acoustic soundproofing structural integrity assessments seismic resistance evaluations flood mitigation plans storm surge barriers levee construction seawall reinforcement breakwater installation harbor dredging channel maintenance buoy placement lighthouse automation radar monitoring sonar mapping bathymetric surveys hydrographic charts nautical almanacs tide tables current predictions weather forecasts marine meteorological models satellite imagery remote sensing data assimilation numerical simulations computational fluid dynamics finite element analysis partial differential equations tensor calculus linear algebra probability theory statistics stochastic processes Markov chains Bayesian inference machine learning artificial neural network deep convolutional recurrent transformer attention heads generative adversarial diffusion transformers large language model tokenization"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-v-x3-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 193,
+  "tok_s": 68.2,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, gripped the iron railing as he ascended the final flight of steps to the lantern room.\n\nBelow him, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t flip a switch. He didn\u2019t press a button. He reached for the heavy brass lever that had worn smooth by his palm over the years. With a grunt that matched the strain of his aging muscles, he pulled it down.\n\n*Clunk.*\n\nThe"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 181,
+  "tok_s": 99.4,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 76.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 105,
+  "tok_s": 65.2,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 147,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 103,
+  "tok_s": 45.7,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to that date and time for context because you asked me about a lighthouse keeper but gave no specifics beyond \u201cshort story,\u201d so let\u2019s set it: isolated coastal station off the Oregon coast during winter storm season, 1987.\n\nMy name is Elias Vane (I made this fictional; real keepers often kept journals signed with initials only), and I\u2019m writing from memory because my logbook got waterlogged when a rogue wave breached lower storage room at dawn\u2014so here\u2019s what happened:<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 109,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 56,
+  "tok_s": 42.1,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they settled into silence with every storm break out of three hundred days here being storms that broke something else again but mostly kept me warm because I burned coal and wood whenever we had any left over from last winter when even then if you\u2019re lucky enough to get through it all without freezing yourself so hard your own thoughts become ice too\n\nBut Elias knew better than anyone: fear was useful only until dawn<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 92,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 48,
+  "tok_s": 43.9,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nHippopotamus          \nRhinoceros            \nCamel              \nCheetah             \nLeopard           \nPanther         \nJaguar                \nBison               \nElk                 \nDeer                  \nFox                   \nWolf                    \nBear                     \nLlama                       \nDonkey                        \nGoat                      \nSheep                         \nCow                           \nPig                          \nRabbit                            \nCat                              \nDog<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 14,
+  "tok_s": 61.4,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 53,
+  "drafts_offered": 134,
+  "tok_s": 51.9,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story we tell ourselves through our relationship** between landmasses separated by vast expansiveness that defines both isolation from other worlds (and thus their own unique evolutionary paths) as well than connection across them via currents wind patterns tides depth pressure temperature salinity biodiversity geology climate change economic trade routes ecological systems cultural exchange political boundaries scientific discovery technological advancement environmental degradation conservation efforts maritime archaeography historical cartographic evolution legal frameworks international law enforcement mechanisms insurance liability regimes regulatory compliance standards safety protocols navigation aids communication networks emergency response strategies search-and-rescue operations salvage rights ownership disputes heritage protection zones sustainable harvesting practices aquaculture development infrastructure engineering projects coastal management policies urban planning regulations zoning ordinances building codes architectural design guidelines aesthetic considerations functional requirements accessibility provisions energy efficiency measures thermal insulation techniques moisture control methods ventilation optimization lighting solutions acoustic soundproofing structural integrity assessments seismic resistance evaluations flood mitigation plans storm surge barriers levee construction seawall reinforcement breakwater installation harbor dredging channel maintenance buoy placement lighthouse automation radar monitoring sonar mapping bathymetric surveys hydrographic charts nautical almanacs tide tables current predictions weather forecasts marine meteorological models satellite imagery remote sensing data assimilation numerical simulations computational fluid dynamics finite element analysis partial differential equations tensor calculus linear algebra probability theory statistics stochastic processes Markov chains Bayesian inference machine learning artificial neural network deep convolutional recurrent transformer attention heads generative adversarial diffusion transformers large language model tokenization"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-x2-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 91,
+  "drafts_offered": 190,
+  "tok_s": 74.0,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, knew every bolt, every lever, and every creak in the iron bones of the tower.\n\nFor forty years, Elias had been the guardian of the Point. Not just of the light, but of the silence between the waves.\n\nIt was November. The sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that shook the floorboards beneath Elias\u2019s boots.\n\nHe climbed the stairs. One hundred and twelve steps. He counted them by feel, his left hand trailing along the cold iron railing. At the top, the air was"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 137,
+  "drafts_offered": 178,
+  "tok_s": 103.4,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 91.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 70.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 38,
+  "drafts_offered": 108,
+  "tok_s": 65.4,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 57,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 61,
+  "tok_s": 57.1,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach top room where his world began again.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 58,
+  "repeats": 0,
+  "drafts_accepted": 14,
+  "drafts_offered": 54,
+  "tok_s": 52.9,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in stone and steel, carrying with them spray from three miles out into churning black water that no longer resembled an ocean so much as abyssal chaos waiting for its prey to surface beneath waves of salt.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 47,
+  "tok_s": 51.7,
+  "text": "Lion  \nTiger  \n\nElephant\\nGiraffe\\Zebra\\rHippopotamus\\tRhinoceros\\vCrocodile\\fAlligator\\bSnake\\aSpider\\u2615CoffeeCup\u2615\ufe0fTeaTime\ud83d\udc09DragonFishWhaleDolph<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 12,
+  "tok_s": 73.5,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 43,
+  "drafts_offered": 126,
+  "tok_s": 63.4,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story told by waves**. It serves as both witness (through sedimentary layers)and participant( through currents that carry messages across continents). Its \"history\" spans geological epochs where life first emerged from primordial pools into complex ecosystems today\u2014and cultural histories shaped entirely around its roleas connector between distant lands or barrier separating them forever depending on perspective taken at different times throughout millennia since ancient civilizations began building boats outof logs carved using stone tools until modern container ships carrying billions tons cargo daily cross Atlantic routes once navigated onlyby stars alone under moonlight during nights when sailors feared monsters lurking beneath hulls made fragile against storms they could barely predict even now despite advanced meteorology systems developed over past century after industrial revolution transformed maritime trade patterns globally due economic pressures driving demand for faster logistics chains linking factories producing electronics manufactured inside cities located far inland near rivers feeding estuaries opening directly onto open seas connected via canals engineered centuries ago like Suez Canal built originallyfor military purposes later adapted commercial use post-World War II era reshaping global supply networks dramatically because geopolitical shifts altered which ports remained accessible politically versus physically blocked temporarily then reopened permanently following treaty agreements signed among nations agreeing upon shared responsibilities regarding environmental stewardship including coral reef restoration efforts funded jointly international organizations coordinating scientific research conducted aboard vessels specifically designed deep-sea exploration equipped robotic submersibles capable descending trenches deeper than any previous technology allowed"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-x3-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 98,
+  "drafts_offered": 166,
+  "tok_s": 54.4,
+  "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 137,
+  "drafts_offered": 163,
+  "tok_s": 70.7,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 56.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 57.6,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 47,
+  "tok_s": 39.7,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBuffalo     \n\nCat          \nCheetah      \n\nChimpanzee \nClam         \n\n\nCrocodile   \n\n\n\nDolphin      \n\n\n\n\nDonkey       <|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 57,
+  "repeats": 0,
+  "drafts_accepted": 12,
+  "drafts_offered": 58,
+  "tok_s": 39.9,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach topmost gallery where his world began.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 30,
+  "drafts_offered": 111,
+  "tok_s": 34.7,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his stone cottage like an angry ghost trying to get inside, yet finding no door left unlocked because Elias never let one stay that way for more than three seconds after entering or leaving during storm season. He had been doing this since 1987\u2014since before he even knew why\u2014and now it was muscle memory so deep into his bones and fingers\u2019 calloused skin texture patterns worn thin by years of gripping the same iron railing at midnight watch that if you asked him what time exactly, down to quarter-hour increments when fog rolled in thick enough where light beams bent around itself like old rope coiling tight on a winch drum spooling out slowly from below deck floor level storage locker number seven under stairs leading"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 96,
+  "repeats": 0,
+  "drafts_accepted": 31,
+  "drafts_offered": 48,
+  "tok_s": 42.3,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nCheetah         \nHippopotamus          \nRhinoceros            \nLeopard             \nJaguar              \nWolf                \nFox                 \nBear                  \nDeer                    \nOtter                   \nBeaver                      \nBadger                        \nMongoose                          \nMeerkat                           \nChimpanzee                            \nGorilla                             \nOrangetutan                              \nAntelope                                \nBison                                   \nElk                                     <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 43.0,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 127,
+  "tok_s": 37.6,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\u201d \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \u201cthe Sea\u201d or \"Ocean\" (terms often used interchangeably but distinct geographically) was never just an element within our planet\u2019s geography\u2014it became its engine itself\u2014a living system whose existence predates humanity by billions years yet continues shaping us today through climate regulation oxygen production food supply trade routes cultural identity ecological resilience economic development political boundaries scientific discovery technological innovation environmental crisis conservation efforts historical memory philosophical reflection poetic metaphor musical composition cinematic storytelling literary narrative visual art architectural inspiration culinary tradition spiritual symbolism psychological influence sociological transformation educational foundation medical necessity agricultural dependency industrial revolution urban planning infrastructural engineering legal framework regulatory policy international cooperation global governance planetary stewardship existential risk mitigation sustainable practice ethical consideration moral imperative social justice democratic participation civic engagement public health community welfare individual responsibility personal accountability collective action shared experience common understanding mutual respect reciprocal trust collaborative effort cooperative endeavor joint venture unified mission integrated strategy coordinated plan synchronized schedule aligned timeline concurrent execution simultaneous operation parallel processing sequential progression iterative refinement recursive improvement continuous adaptation dynamic adjustment flexible response adaptive behavior responsive conduct proactive initiative anticipatory foresight predictive modeling probabilistic assessment statistical analysis empirical observation experimental verification theoretical hypothesis conceptual abstraction logical deduction rational inference critical evaluation analytical reasoning systematic investigation methodical inquiry structured exploration comprehensive examination detailed scrutiny meticulous inspection thorough review exhaustive audit complete survey full census total enumeration absolute"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-dev-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 59.2,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 81.6,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.9,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 49.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 44.9,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 44.1,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 49.5,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 65.2,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.5,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-rel-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 58.4,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 80.6,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 75.2,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.2,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 48.5,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 46.4,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 43.3,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 49.8,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.5,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.6,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-same24-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 58.4,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 80.5,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 76.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 49.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 46.7,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 44.3,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 50.0,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.0,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.1,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/pp-dev-3090.json

@@ -0,0 +1,72 @@
+{
+ "label": "dev-3090",
+ "extra": [],
+ "gpus": "2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 3424.8,
+   "pp_tok_s": 796.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 8617.9,
+   "pp_tok_s": 972.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 12820.1,
+   "pp_tok_s": 1253.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 22901.1,
+   "pp_tok_s": 1211.4,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 5657.8,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 238.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 71.7,
+   "hit_rate": 0.9266798230257504,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, punctuated by whitecaps that looked like teeth.\n\nHe climbed the spiral stairs, his knees protesting with every step. The iron rungs were cold, slick with condensation that had leaked through the hairline cracks in the masonry. At the top, the lantern room hummed with the low, steady thrum of the rotating lens. It was a sound Elias had learned to sleep with, a heartbeat of the house.\n\nHe checked the clock. 11:42 PM. The light swept across the water, a slow, hypnotic arc. *Flash. Darkness. Flash. Darkness.*\n\nFor a moment, the beam caught something that wasn\u2019t water. A shape"
+  },
+  {
+   "name": "code",
+   "tok_s": 92.9,
+   "hit_rate": 0.8636411400016346,
+   "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header row\n    and values from the corresponding row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n```<|im_end|>"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 17.31 GiB free, 700 MiB reserved -> 6704 slots",
+  "strata generate: expert cache 8730 slots, 16.62 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2315 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 86.6% (4986 hits / 5756 lookups)",
+  "strata serve: decode expert cache hit rate: 87.1% (3757 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 88.7% (3827 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 92.1% (3975 hits / 4318 lookups)",
+  "strata serve: decode expert cache hit rate: 95.5% (4124 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 94.3% (25793 hits / 27360 lookups)",
+  "strata serve: decode expert cache hit rate: 94.6% (4088 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 92.7% (116454 hits / 125668 lookups)",
+  "strata serve: decode expert cache hit rate: 86.4% (73970 hits / 85649 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-dev-5080.json

@@ -0,0 +1,72 @@
+{
+ "label": "dev-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2129.8,
+   "pp_tok_s": 1279.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5057.2,
+   "pp_tok_s": 1656.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7964.7,
+   "pp_tok_s": 2017.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14053.6,
+   "pp_tok_s": 1974.1,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3366.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 197.3,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.1,
+   "hit_rate": 0.8418683557684931,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 88.0,
+   "hit_rate": 0.7011385968111186,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 72.5% (2539 hits / 3504 lookups)",
+  "strata serve: decode expert cache hit rate: 77.3% (2817 hits / 3642 lookups)",
+  "strata serve: decode expert cache hit rate: 75.3% (16500 hits / 21916 lookups)",
+  "strata serve: decode expert cache hit rate: 80.0% (2978 hits / 3723 lookups)",
+  "strata serve: decode expert cache hit rate: 84.2% (98752 hits / 117301 lookups)",
+  "strata serve: decode expert cache hit rate: 70.1% (69215 hits / 98718 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k18.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k18",
+ "extra": [
+  "--layer-split",
+  "18"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 10383.7,
+   "pp_tok_s": 1547.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 16371.9,
+   "pp_tok_s": 1694.5,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4308.0,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 239.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 77.2,
+   "reps": [
+    74.6,
+    78.1,
+    79.0
+   ],
+   "hit_rate": 0.9985378787878788,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His knees ached, not from age, but from the constant vibration of the floor beneath his feet. He reached the lantern room, the heart of the lighthouse, and checked the clockwork mechanism. The great lens, a prism of Fresnel glass, rotated with a steady, hypnotic *whirrr-click, whirrr-click*. It was the only sound that mattered. The light swept out into the black void, cutting through"
+  },
+  {
+   "name": "code",
+   "tok_s": 103.8,
+   "reps": [
+    97.6,
+    106.4,
+    107.3
+   ],
+   "hit_rate": 0.9984863945578232,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 524 windows; per window: wait for the GPU 6.456 ms, pool + plan 0.181 ms, host staging 0.581 ms, commit 0.206 ms",
+  "strata serve: stage 1: 524 windows; per window: wait for the GPU 14.872 ms, pool + plan 0.501 ms, host staging 0.000 ms, commit 0.387 ms",
+  "strata serve: stage 0: 593 windows; per window: wait for the GPU 6.479 ms, pool + plan 0.185 ms, host staging 0.515 ms, commit 0.207 ms",
+  "strata serve: stage 1: 593 windows; per window: wait for the GPU 15.053 ms, pool + plan 0.496 ms, host staging 0.000 ms, commit 0.390 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 18)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4608 slots, 8.24 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 18-47, expert cache 7152 slots (14.14 GiB), 7152 of its 7680 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-17",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-17 (CUDA0), 18-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.7% (6159 hits / 6240 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (3796 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (3811 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (27723 hits / 27840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (3807 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (142002 hits / 143040 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (139882 hits / 140160 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (131807 hits / 132000 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (119278 hits / 120480 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (117410 hits / 117600 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (117422 hits / 117600 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k19.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k19",
+ "extra": [
+  "--layer-split",
+  "19"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9875.2,
+   "pp_tok_s": 1626.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15498.0,
+   "pp_tok_s": 1790.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4085.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 238.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 79.3,
+   "reps": [
+    74.8,
+    85.6,
+    77.4
+   ],
+   "hit_rate": 0.9977757017660928,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias stood at the top of the spiral staircase, his hand resting on the cold iron railing. He was not afraid. Fear was for the tourists who came in summer, wide-eyed and clutching their cameras. Elias was a keeper. He was a guardian.\n\nHe checked the logbook. The entry for the night was simple: *Wind, 60 knots. Visibility, zero. Light, operational.*\n\nHe walked down the stairs, his boots heavy on the metal grating. The smell of salt"
+  },
+  {
+   "name": "code",
+   "tok_s": 107.4,
+   "reps": [
+    98.9,
+    111.3,
+    112.1
+   ],
+   "hit_rate": 0.9986286196529353,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header row\n    and values from the corresponding row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 533 windows; per window: wait for the GPU 6.226 ms, pool + plan 0.294 ms, host staging 0.561 ms, commit 0.225 ms",
+  "strata serve: stage 1: 533 windows; per window: wait for the GPU 14.156 ms, pool + plan 0.373 ms, host staging 0.000 ms, commit 0.371 ms",
+  "strata serve: stage 0: 599 windows; per window: wait for the GPU 6.314 ms, pool + plan 0.292 ms, host staging 0.501 ms, commit 0.225 ms",
+  "strata serve: stage 1: 599 windows; per window: wait for the GPU 14.366 ms, pool + plan 0.375 ms, host staging 0.000 ms, commit 0.374 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 19)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4606 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7160 slots (14.14 GiB), 7160 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.5% (6141 hits / 6234 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (3797 hits / 3836 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3814 hits / 3837 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (28659 hits / 28795 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3814 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (141378 hits / 142508 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (124963 hits / 125273 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (134574 hits / 134874 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (121132 hits / 122322 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (114073 hits / 114235 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (113598 hits / 113754 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k20.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k20",
+ "extra": [
+  "--layer-split",
+  "20"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9044.3,
+   "pp_tok_s": 1776.4,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13928.7,
+   "pp_tok_s": 1991.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3743.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 236.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 76.4,
+   "reps": [
+    74.2,
+    78.4,
+    76.7
+   ],
+   "hit_rate": 0.9979911611088791,
+   "text": "The wind didn\u2019t just blow; it screamed. It clawed at the granite base of the tower, trying to pry it from the bedrock, and it howled through the narrow slits of the lantern room glass.\n\nElias Thorne was seventy years old, but he moved with the grace of a man who had spent fifty years listening to the sea. He knew the rhythm of the waves against the rocks below\u2014the *thud, thud, crash*\u2014and he knew when that rhythm changed. When the sea grew too quiet, or too violent, he knew a storm was coming, or a ship was in trouble.\n\nTonight, the storm was a Category 4. The rain lashed against the windows like gravel thrown by a giant. Elias stood in the center of the lantern room, his hand resting on the cold brass railing. He wasn\u2019t afraid. He was a guardian.\n\nHe checked the clockwork mechanism. It was old, a relic from 1"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.9,
+   "reps": [
+    103.2,
+    106.8,
+    107.7
+   ],
+   "hit_rate": 0.9981849807154201,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 537 windows; per window: wait for the GPU 6.685 ms, pool + plan 0.403 ms, host staging 0.593 ms, commit 0.215 ms",
+  "strata serve: stage 1: 537 windows; per window: wait for the GPU 13.928 ms, pool + plan 0.282 ms, host staging 0.000 ms, commit 0.359 ms",
+  "strata serve: stage 0: 606 windows; per window: wait for the GPU 6.764 ms, pool + plan 0.399 ms, host staging 0.528 ms, commit 0.216 ms",
+  "strata serve: stage 1: 606 windows; per window: wait for the GPU 14.108 ms, pool + plan 0.288 ms, host staging 0.000 ms, commit 0.362 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 20)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 7142 slots (14.14 GiB), 7142 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.9% (6162 hits / 6230 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (3803 hits / 3835 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3810 hits / 3836 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27185 hits / 27330 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3807 hits / 3835 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (139825 hits / 140938 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (142122 hits / 142524 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (146556 hits / 146851 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (118026 hits / 119233 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118729 hits / 118997 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118791 hits / 119007 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k21.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k21",
+ "extra": [
+  "--layer-split",
+  "21"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8880.5,
+   "pp_tok_s": 1809.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13654.6,
+   "pp_tok_s": 2031.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3680.6,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 235.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 77.4,
+   "reps": [
+    80.2,
+    80.8,
+    71.2
+   ],
+   "hit_rate": 0.9973622550922132,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from rope burns and broken glass, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, ensuring the rotation was smooth, silent, and steady. *Tick. Tick. Tick.* The sound was the heartbeat of the coast.\n\n"
+  },
+  {
+   "name": "code",
+   "tok_s": 106.7,
+   "reps": [
+    100.6,
+    110.5,
+    109.0
+   ],
+   "hit_rate": 0.998218711927068,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 527 windows; per window: wait for the GPU 6.934 ms, pool + plan 0.620 ms, host staging 0.530 ms, commit 0.245 ms",
+  "strata serve: stage 1: 527 windows; per window: wait for the GPU 13.912 ms, pool + plan 0.365 ms, host staging 0.000 ms, commit 0.359 ms",
+  "strata serve: stage 0: 595 windows; per window: wait for the GPU 7.031 ms, pool + plan 0.595 ms, host staging 0.472 ms, commit 0.245 ms",
+  "strata serve: stage 1: 595 windows; per window: wait for the GPU 14.051 ms, pool + plan 0.361 ms, host staging 0.000 ms, commit 0.361 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 21)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4627 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 21-47, expert cache 6912 slots (13.78 GiB), 6912 of its 6912 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-20",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-20 (CUDA0), 21-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.7% (5672 hits / 5748 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4287 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4287 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (29024 hits / 29229 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4316 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (131499 hits / 133085 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (130974 hits / 131451 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (138767 hits / 139134 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (121454 hits / 122932 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (116809 hits / 117079 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118803 hits / 119015 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k22.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k22",
+ "extra": [
+  "--layer-split",
+  "22"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8730.9,
+   "pp_tok_s": 1840.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13375.4,
+   "pp_tok_s": 2074.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3611.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 234.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 80.2,
+   "reps": [
+    78.4,
+    80.9,
+    81.5
+   ],
+   "hit_rate": 0.9977764197412695,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and he had been the keeper of the Blackwood Point Light for forty of those years. The other thirty-three had been spent in a small, drafty cottage on the mainland, waiting for the day when the sea would finally claim him.\n\nTonight, the storm was a Category 4. The waves weren\u2019t just crashing; they were climbing. They reached the gallery railing, ten feet above the base of the tower, spraying a fine mist of brine that coated Elias\u2019s glasses.\n\nHe climbed the spiral stairs, his knees protesting with every step. The iron steps were cold, slick with condensation. At the top, in the lantern room, the air was thin and sharp. The great Fresnel lens, a cathedral of glass and brass, stood silent. It was time to wake it.\n\n"
+  },
+  {
+   "name": "code",
+   "tok_s": 106.4,
+   "reps": [
+    100.7,
+    110.3,
+    108.1
+   ],
+   "hit_rate": 0.9971798370756316,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 529 windows; per window: wait for the GPU 7.229 ms, pool + plan 0.561 ms, host staging 0.529 ms, commit 0.233 ms",
+  "strata serve: stage 1: 529 windows; per window: wait for the GPU 13.058 ms, pool + plan 0.253 ms, host staging 0.000 ms, commit 0.334 ms",
+  "strata serve: stage 0: 598 windows; per window: wait for the GPU 7.325 ms, pool + plan 0.552 ms, host staging 0.469 ms, commit 0.235 ms",
+  "strata serve: stage 1: 598 windows; per window: wait for the GPU 13.238 ms, pool + plan 0.260 ms, host staging 0.000 ms, commit 0.337 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 22)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.7% (4256 hits / 4310 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (27008 hits / 27289 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (4260 hits / 4313 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (135183 hits / 137274 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (136132 hits / 136708 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (135515 hits / 135817 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (121287 hits / 123280 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (117658 hits / 118026 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (120574 hits / 120915 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k23.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k23",
+ "extra": [
+  "--layer-split",
+  "23"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8742.6,
+   "pp_tok_s": 1837.7,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13084.2,
+   "pp_tok_s": 2120.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3566.1,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 238.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 81.0,
+   "reps": [
+    76.8,
+    85.2,
+    81.1
+   ],
+   "hit_rate": 0.9961302681992337,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, swallowing the horizon, throwing white foam against the granite base like a tantrum.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his wool trousers, the fabric stiff with salt and grease. He checked the clockwork mechanism that turned the lens. It was a beast of a thing, a Fresnel lens of eight feet in diameter, a cathedral of glass and crystal that focused the beam of the lamp into a single, piercing needle of light.\n\nHe wound the weights. *Click, click, click.* The gears engaged. The massive lens began its slow, hypnotic rotation.\n\nFor a moment, the"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.4,
+   "reps": [
+    105.1,
+    109.0,
+    111.1
+   ],
+   "hit_rate": 0.9961698810290479,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 537 windows; per window: wait for the GPU 7.423 ms, pool + plan 0.953 ms, host staging 0.542 ms, commit 0.237 ms",
+  "strata serve: stage 1: 537 windows; per window: wait for the GPU 12.223 ms, pool + plan 0.236 ms, host staging 0.000 ms, commit 0.324 ms",
+  "strata serve: stage 0: 606 windows; per window: wait for the GPU 7.519 ms, pool + plan 0.904 ms, host staging 0.483 ms, commit 0.239 ms",
+  "strata serve: stage 1: 606 windows; per window: wait for the GPU 12.395 ms, pool + plan 0.242 ms, host staging 0.000 ms, commit 0.326 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 23)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4609 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 23-47, expert cache 6400 slots (12.82 GiB), 6400 of its 6400 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-22",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-22 (CUDA0), 23-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 96.8% (5528 hits / 5709 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (4235 hits / 4303 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4260 hits / 4308 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (30643 hits / 31089 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (4248 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 98.3% (134776 hits / 137143 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (133024 hits / 133766 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (129995 hits / 130500 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (115782 hits / 118287 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (117953 hits / 118464 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (117560 hits / 118012 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k24.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k24",
+ "extra": [
+  "--layer-split",
+  "24"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8122.5,
+   "pp_tok_s": 1978.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12285.8,
+   "pp_tok_s": 2258.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3407.0,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 234.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.5,
+   "reps": [
+    84.9,
+    83.3,
+    82.3
+   ],
+   "hit_rate": 0.9949381327334084,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined Elias\u2019s life for thirty years. He was not merely a keeper; he was a guardian of silence.\n\nHe checked the clock. 11:42 PM. The rotation of the great Fresnel lens was due. With a practiced hand, he engaged the brake on the clockwork mechanism. The heavy brass gears groaned, then began their slow, rhythmic turn. *Tick. Tick. Tick.* It was a heartbeat"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.5,
+   "reps": [
+    104.9,
+    111.0,
+    112.6
+   ],
+   "hit_rate": 0.9953796671318199,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 519 windows; per window: wait for the GPU 7.835 ms, pool + plan 0.767 ms, host staging 0.557 ms, commit 0.244 ms",
+  "strata serve: stage 1: 519 windows; per window: wait for the GPU 11.922 ms, pool + plan 0.232 ms, host staging 0.000 ms, commit 0.325 ms",
+  "strata serve: stage 0: 587 windows; per window: wait for the GPU 7.943 ms, pool + plan 0.744 ms, host staging 0.497 ms, commit 0.246 ms",
+  "strata serve: stage 1: 587 windows; per window: wait for the GPU 12.073 ms, pool + plan 0.238 ms, host staging 0.000 ms, commit 0.326 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 24)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 24-47, expert cache 6144 slots (12.35 GiB), 6144 of its 6144 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-23",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-23 (CUDA0), 24-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 95.9% (5916 hits / 6171 lookups)",
+  "strata serve: decode expert cache hit rate: 97.4% (3717 hits / 3815 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (3761 hits / 3823 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (27138 hits / 27664 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (3742 hits / 3826 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (126114 hits / 128875 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (133626 hits / 134658 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (132675 hits / 133350 lookups)",
+  "strata serve: decode expert cache hit rate: 96.8% (113593 hits / 117399 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (118372 hits / 118949 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (116981 hits / 117524 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k26.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k26",
+ "extra": [
+  "--layer-split",
+  "26"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 7880.9,
+   "pp_tok_s": 2038.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 11771.9,
+   "pp_tok_s": 2356.7,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3323.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 231.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.8,
+   "reps": [
+    85.9,
+    83.1,
+    82.4
+   ],
+   "hit_rate": 0.9933966961830739,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and green.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined Elias\u2019s life for thirty years. He was a keeper of the light, but more importantly, he was a keeper of time.\n\nHe checked his pocket watch. 11:42 PM. The rotation of the great Fresnel lens was due. With a practiced, rhythmic *clunk-clunk-clunk*, he wound the clockwork mechanism. The massive lens began its slow, hypnotic turn, sweeping its beam across"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.7,
+   "reps": [
+    104.4,
+    112.7,
+    111.9
+   ],
+   "hit_rate": 0.9938670271352759,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 519 windows; per window: wait for the GPU 8.376 ms, pool + plan 1.030 ms, host staging 0.566 ms, commit 0.273 ms",
+  "strata serve: stage 1: 519 windows; per window: wait for the GPU 10.886 ms, pool + plan 0.214 ms, host staging 0.000 ms, commit 0.297 ms",
+  "strata serve: stage 0: 588 windows; per window: wait for the GPU 8.506 ms, pool + plan 0.985 ms, host staging 0.505 ms, commit 0.274 ms",
+  "strata serve: stage 1: 588 windows; per window: wait for the GPU 11.036 ms, pool + plan 0.219 ms, host staging 0.000 ms, commit 0.298 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 26)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4568 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 26-47, expert cache 5632 slots (11.32 GiB), 5632 of its 5632 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-25",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-25 (CUDA0), 26-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 92.8% (5259 hits / 5665 lookups)",
+  "strata serve: decode expert cache hit rate: 95.2% (4056 hits / 4259 lookups)",
+  "strata serve: decode expert cache hit rate: 97.2% (4164 hits / 4285 lookups)",
+  "strata serve: decode expert cache hit rate: 97.1% (27690 hits / 28510 lookups)",
+  "strata serve: decode expert cache hit rate: 97.5% (4184 hits / 4293 lookups)",
+  "strata serve: decode expert cache hit rate: 96.9% (121839 hits / 125724 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (132165 hits / 133648 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (137652 hits / 138567 lookups)",
+  "strata serve: decode expert cache hit rate: 95.3% (111307 hits / 116822 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (118436 hits / 119329 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (119109 hits / 119844 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k28.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k28",
+ "extra": [
+  "--layer-split",
+  "28"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 7989.3,
+   "pp_tok_s": 2010.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12132.9,
+   "pp_tok_s": 2286.6,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3433.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 228.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.5,
+   "reps": [
+    83.7,
+    82.4,
+    84.5
+   ],
+   "hit_rate": 0.9900441970361987,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His knees ached, not from age, but from the constant vibration of the floor beneath his feet. He reached the lantern room, the heart of the lighthouse, and checked the clockwork mechanism. The great lens, a prism of Fresnel glass, rotated with a steady, hypnotic *whirrr-click, whirrr-click*. It was the only sound that mattered. The light swept out into the black void, cutting through"
+  },
+  {
+   "name": "code",
+   "tok_s": 112.5,
+   "reps": [
+    101.4,
+    112.1,
+    124.1
+   ],
+   "hit_rate": 0.9896636587366694,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n```<|im_end|>"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 527 windows; per window: wait for the GPU 8.765 ms, pool + plan 1.401 ms, host staging 0.593 ms, commit 0.281 ms",
+  "strata serve: stage 1: 527 windows; per window: wait for the GPU 9.893 ms, pool + plan 0.195 ms, host staging 0.000 ms, commit 0.276 ms",
+  "strata serve: stage 0: 578 windows; per window: wait for the GPU 8.908 ms, pool + plan 1.359 ms, host staging 0.559 ms, commit 0.284 ms",
+  "strata serve: stage 1: 578 windows; per window: wait for the GPU 10.025 ms, pool + plan 0.200 ms, host staging 0.000 ms, commit 0.278 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 28)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4537 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 28-47, expert cache 5120 slots (10.30 GiB), 5120 of its 5120 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-27",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-27 (CUDA0), 28-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 88.8% (5405 hits / 6086 lookups)",
+  "strata serve: decode expert cache hit rate: 93.8% (3539 hits / 3773 lookups)",
+  "strata serve: decode expert cache hit rate: 94.2% (3560 hits / 3781 lookups)",
+  "strata serve: decode expert cache hit rate: 95.7% (26702 hits / 27899 lookups)",
+  "strata serve: decode expert cache hit rate: 95.9% (3647 hits / 3804 lookups)",
+  "strata serve: decode expert cache hit rate: 95.7% (122618 hits / 128160 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (135035 hits / 136884 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (129476 hits / 130778 lookups)",
+  "strata serve: decode expert cache hit rate: 93.9% (111037 hits / 118290 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (118829 hits / 120140 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (90480 hits / 91425 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k30.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k30",
+ "extra": [
+  "--layer-split",
+  "30"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8301.3,
+   "pp_tok_s": 1935.4,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12507.6,
+   "pp_tok_s": 2218.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3448.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 229.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 81.1,
+   "reps": [
+    78.0,
+    84.4,
+    81.0
+   ],
+   "hit_rate": 0.9798126997277784,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias stood at the top of the spiral staircase, his hand resting on the cold iron railing. He was not afraid. Fear was for the tourists who came in summer, wide-eyed and clutching their cameras. Elias was a keeper. He was a guardian.\n\nHe checked the logbook. The entry for the night was simple: *Wind, 60 knots. Visibility, zero. Light, operational.*\n\nHe walked down the stairs, his boots heavy on the metal grating. The smell of salt"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.6,
+   "reps": [
+    100.7,
+    114.1,
+    114.1
+   ],
+   "hit_rate": 0.9825370343977666,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 540 windows; per window: wait for the GPU 9.206 ms, pool + plan 1.794 ms, host staging 0.525 ms, commit 0.298 ms",
+  "strata serve: stage 1: 540 windows; per window: wait for the GPU 8.767 ms, pool + plan 0.173 ms, host staging 0.000 ms, commit 0.242 ms",
+  "strata serve: stage 0: 608 windows; per window: wait for the GPU 9.390 ms, pool + plan 1.733 ms, host staging 0.471 ms, commit 0.301 ms",
+  "strata serve: stage 1: 608 windows; per window: wait for the GPU 8.910 ms, pool + plan 0.178 ms, host staging 0.000 ms, commit 0.244 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 30)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4479 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 30-47, expert cache 4608 slots (9.19 GiB), 4608 of its 4608 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-29",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-29 (CUDA0), 30-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 84.6% (5099 hits / 6028 lookups)",
+  "strata serve: decode expert cache hit rate: 90.7% (3403 hits / 3751 lookups)",
+  "strata serve: decode expert cache hit rate: 91.1% (3415 hits / 3747 lookups)",
+  "strata serve: decode expert cache hit rate: 94.1% (26108 hits / 27737 lookups)",
+  "strata serve: decode expert cache hit rate: 93.4% (3534 hits / 3784 lookups)",
+  "strata serve: decode expert cache hit rate: 94.6% (127532 hits / 134866 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (135828 hits / 137786 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (132455 hits / 135184 lookups)",
+  "strata serve: decode expert cache hit rate: 91.0% (106351 hits / 116917 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (115791 hits / 118246 lookups)",
+  "strata serve: decode expert cache hit rate: 98.3% (117198 hits / 119281 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k32.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k32",
+ "extra": [
+  "--layer-split",
+  "32"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8837.1,
+   "pp_tok_s": 1818.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13449.8,
+   "pp_tok_s": 2062.7,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3565.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 227.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.0,
+   "reps": [
+    75.0,
+    86.1,
+    87.8
+   ],
+   "hit_rate": 0.9796934433020751,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.7,
+   "reps": [
+    98.8,
+    112.2,
+    115.1
+   ],
+   "hit_rate": 0.9777186846317358,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 533 windows; per window: wait for the GPU 9.839 ms, pool + plan 2.192 ms, host staging 0.585 ms, commit 0.318 ms",
+  "strata serve: stage 1: 533 windows; per window: wait for the GPU 8.014 ms, pool + plan 0.156 ms, host staging 0.000 ms, commit 0.231 ms",
+  "strata serve: stage 0: 602 windows; per window: wait for the GPU 10.014 ms, pool + plan 2.111 ms, host staging 0.520 ms, commit 0.321 ms",
+  "strata serve: stage 1: 602 windows; per window: wait for the GPU 8.094 ms, pool + plan 0.160 ms, host staging 0.000 ms, commit 0.232 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 32)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4488 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 32-47, expert cache 4096 slots (8.26 GiB), 4096 of its 4096 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-31",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-31 (CUDA0), 32-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 81.3% (4889 hits / 6010 lookups)",
+  "strata serve: decode expert cache hit rate: 87.7% (3272 hits / 3733 lookups)",
+  "strata serve: decode expert cache hit rate: 87.9% (3280 hits / 3733 lookups)",
+  "strata serve: decode expert cache hit rate: 92.3% (24238 hits / 26254 lookups)",
+  "strata serve: decode expert cache hit rate: 91.5% (3457 hits / 3778 lookups)",
+  "strata serve: decode expert cache hit rate: 93.7% (133254 hits / 142240 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (130112 hits / 132812 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (130069 hits / 132765 lookups)",
+  "strata serve: decode expert cache hit rate: 89.1% (105263 hits / 118190 lookups)",
+  "strata serve: decode expert cache hit rate: 97.2% (116970 hits / 120309 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (115450 hits / 118081 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-single.json

@@ -0,0 +1,76 @@
+{
+ "label": "ks-single",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9310.4,
+   "pp_tok_s": 1725.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14085.4,
+   "pp_tok_s": 1969.6,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3388.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 200.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 86.7,
+   "reps": [
+    78.2,
+    91.4,
+    90.4
+   ],
+   "hit_rate": 0.921882148773369,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and his hands were a map of the coastline he had guarded for fifty years. The knuckles were swollen, the skin cracked like dry earth, but they moved with a precision that no machine could replicate.\n\nHe stood on the gallery of the *Harrow Point* lighthouse, the iron railing cold against his palms. Below, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t look at the gauge. He didn\u2019t check the fuel levels. He knew. He always knew. The rhythm of the ocean, the pressure of the air, the subtle vibration"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.2,
+   "reps": [
+    87.8,
+    113.5,
+    114.2
+   ],
+   "hit_rate": 0.9109811738648947,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 382 windows; per window: wait for the GPU 15.155 ms, pool + plan 4.331 ms, host staging 0.656 ms, commit 0.458 ms",
+  "strata serve: stage 0: 455 windows; per window: wait for the GPU 15.790 ms, pool + plan 5.552 ms, host staging 0.688 ms, commit 0.467 ms",
+  "strata serve: stage 0: 524 windows; per window: wait for the GPU 16.064 ms, pool + plan 5.529 ms, host staging 0.606 ms, commit 0.472 ms",
+  "strata serve: stage 0: 593 windows; per window: wait for the GPU 16.305 ms, pool + plan 5.462 ms, host staging 0.540 ms, commit 0.475 ms"
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 67.8% (2299 hits / 3391 lookups)",
+  "strata serve: decode expert cache hit rate: 68.5% (2345 hits / 3424 lookups)",
+  "strata serve: decode expert cache hit rate: 75.6% (16605 hits / 21959 lookups)",
+  "strata serve: decode expert cache hit rate: 73.3% (2608 hits / 3558 lookups)",
+  "strata serve: decode expert cache hit rate: 83.7% (104878 hits / 125346 lookups)",
+  "strata serve: decode expert cache hit rate: 92.7% (119739 hits / 129184 lookups)",
+  "strata serve: decode expert cache hit rate: 92.2% (116867 hits / 126770 lookups)",
+  "strata serve: decode expert cache hit rate: 68.1% (64111 hits / 94116 lookups)",
+  "strata serve: decode expert cache hit rate: 89.0% (96370 hits / 108259 lookups)",
+  "strata serve: decode expert cache hit rate: 91.1% (102827 hits / 112875 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-rel-5080.json

@@ -0,0 +1,72 @@
+{
+ "label": "rel-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2132.8,
+   "pp_tok_s": 1278.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5008.9,
+   "pp_tok_s": 1672.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7855.4,
+   "pp_tok_s": 2045.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13833.9,
+   "pp_tok_s": 2005.4,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3298.2,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 197.9,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.2,
+   "hit_rate": 0.8418683557684931,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 88.2,
+   "hit_rate": 0.7011385968111186,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 72.5% (2539 hits / 3504 lookups)",
+  "strata serve: decode expert cache hit rate: 77.3% (2817 hits / 3642 lookups)",
+  "strata serve: decode expert cache hit rate: 75.3% (16500 hits / 21916 lookups)",
+  "strata serve: decode expert cache hit rate: 80.0% (2978 hits / 3723 lookups)",
+  "strata serve: decode expert cache hit rate: 84.2% (98752 hits / 117301 lookups)",
+  "strata serve: decode expert cache hit rate: 70.1% (69215 hits / 98718 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-t75-5080.json

@@ -0,0 +1,78 @@
+{
+ "label": "t75-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2158.8,
+   "pp_tok_s": 1262.7,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5060.0,
+   "pp_tok_s": 1655.9,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3389.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 213.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 80.4,
+   "reps": [
+    80.4
+   ],
+   "hit_rate": 0.855538332528651,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, punctuated by whitecaps that looked like teeth.\n\nHe climbed the spiral stairs, his boots heavy with salt and exhaustion. The iron steps groaned under his weight, a sound that had become so familiar to him that he could sleep to it. But tonight, he wasn\u2019t sleeping. He was working.\n\nThe lens was the heart of the lighthouse. A massive, rotating Fresnel lens, a cathedral of glass and brass that turned with a mechanical hum. Elias checked the clockwork mechanism. It was smooth, rhythmic, a heartbeat in the dark. But the light itself was failing.\n\nHe climbed higher, into the lantern room. The wind battered the glass panes,"
+  },
+  {
+   "name": "code",
+   "tok_s": 87.5,
+   "reps": [
+    87.5
+   ],
+   "hit_rate": 0.7075461363025227,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 39 windows; per window: wait for the GPU 17.252 ms, pool + plan 8.473 ms, host staging 0.611 ms, commit 0.525 ms",
+  "strata serve: stage 0: 47 windows; per window: wait for the GPU 18.018 ms, pool + plan 8.616 ms, host staging 0.607 ms, commit 0.534 ms",
+  "strata serve: stage 0: 169 windows; per window: wait for the GPU 15.223 ms, pool + plan 5.717 ms, host staging 0.688 ms, commit 0.467 ms",
+  "strata serve: stage 0: 243 windows; per window: wait for the GPU 16.500 ms, pool + plan 7.316 ms, host staging 0.742 ms, commit 0.481 ms"
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4294 slots, 8.18 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 75.5% (16316 hits / 21602 lookups)",
+  "strata serve: decode expert cache hit rate: 74.4% (2660 hits / 3573 lookups)",
+  "strata serve: decode expert cache hit rate: 83.4% (92664 hits / 111086 lookups)",
+  "strata serve: decode expert cache hit rate: 72.0% (70946 hits / 98601 lookups)",
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 75.5% (16316 hits / 21602 lookups)",
+  "strata serve: decode expert cache hit rate: 74.4% (2659 hits / 3572 lookups)",
+  "strata serve: decode expert cache hit rate: 85.6% (101004 hits / 118059 lookups)",
+  "strata serve: decode expert cache hit rate: 70.8% (71044 hits / 100409 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x2-auto.json

@@ -0,0 +1,103 @@
+{
+ "label": "v-x2-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2577.0,
+   "pp_tok_s": 1057.8,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 4433.5,
+   "pp_tok_s": 1889.9,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7885.8,
+   "pp_tok_s": 2037.3,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13384.6,
+   "pp_tok_s": 2072.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3667.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 253.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 79.5,
+   "reps": [
+    78.9,
+    79.0,
+    80.7
+   ],
+   "hit_rate": 0.9969903443250137,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from rope burns and broken glass, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, listening for the faint, rhythmic *tick-tick-tick* that signaled the rotation was true. If the rhythm faltered, ships"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.9,
+   "reps": [
+    103.3,
+    113.7,
+    109.5
+   ],
+   "hit_rate": 0.9971795386830036,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 542 windows; per window: wait for the GPU 7.167 ms, pool + plan 0.572 ms, host staging 0.550 ms, commit 0.232 ms",
+  "strata serve: stage 1: 542 windows; per window: wait for the GPU 12.855 ms, pool + plan 0.250 ms, host staging 0.000 ms, commit 0.334 ms",
+  "strata serve: stage 0: 610 windows; per window: wait for the GPU 7.276 ms, pool + plan 0.559 ms, host staging 0.493 ms, commit 0.234 ms",
+  "strata serve: stage 1: 610 windows; per window: wait for the GPU 13.048 ms, pool + plan 0.258 ms, host staging 0.000 ms, commit 0.337 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.30 GiB for experts",
+  "strata generate: layer split auto: CUDA1 82 SMs at 1.86 GHz -> 0.48 ms per layer, 14.14 GiB for experts",
+  "strata generate: layer split auto: K=22 - predicted 19.7 ms per decode window; the caches hold 11298 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (4241 hits / 4306 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4262 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4290 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4319 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (28379 hits / 28721 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (4285 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (136232 hits / 138256 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (135540 hits / 136216 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (136812 hits / 137225 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (118437 hits / 120406 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (112836 hits / 113230 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x2r-auto.json

@@ -0,0 +1,88 @@
+{
+ "label": "v-x2r-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "2,0",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9610.0,
+   "pp_tok_s": 1671.8,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14786.1,
+   "pp_tok_s": 1876.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3850.8,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 245.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 90.3,
+   "reps": [
+    91.0,
+    97.0,
+    82.8
+   ],
+   "hit_rate": 0.993905472636816,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak of the spiral staircase, every draft that slipped under the door, and every shift in the barometric pressure that preceded a storm. He was not merely a worker; he was the heartbeat of the coast.\n\nIt was November, the month of grey seas and shorter days. The fog had rolled in thick, a wall of white so dense that even the beam of the Fresnel lens seemed to struggle against it. Elias stood in the lantern room, his hands roughened by salt and rope, adjusting the rotation mechanism. The great lens, a cathedral of glass and brass, began its slow, rhythmic sweep. *Flash. Dark. Flash. Dark.*\n\nHe checked the log"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.1,
+   "reps": [
+    98.4,
+    109.9,
+    107.0
+   ],
+   "hit_rate": 0.9949900793650793,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 492 windows; per window: wait for the GPU 15.016 ms, pool + plan 0.297 ms, host staging 0.517 ms, commit 0.404 ms",
+  "strata serve: stage 1: 492 windows; per window: wait for the GPU 5.775 ms, pool + plan 0.854 ms, host staging 0.001 ms, commit 0.210 ms",
+  "strata serve: stage 0: 561 windows; per window: wait for the GPU 15.257 ms, pool + plan 0.306 ms, host staging 0.455 ms, commit 0.408 ms",
+  "strata serve: stage 1: 561 windows; per window: wait for the GPU 5.852 ms, pool + plan 0.820 ms, host staging 0.001 ms, commit 0.210 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 57.7 GB/s -> pcie_frac 0.55",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 9.98 GiB free",
+  "strata generate: layer split auto: CUDA0 82 SMs at 1.86 GHz -> 0.48 ms per layer, 16.45 GiB for experts",
+  "strata generate: layer split auto: CUDA1 84 SMs at 2.81 GHz -> 0.31 ms per layer, 5.99 GiB for experts",
+  "strata generate: layer split auto: K=30 - predicted 20.8 ms per decode window; the caches hold 10684 of 12288 profiled pairs (~99.5% of the routed mass)",
+  "strata generate: expert cache auto: 18.62 GiB free, 700 MiB reserved -> 6634 slots",
+  "strata generate: expert cache 7680 slots, 14.23 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 30-47, expert cache 3004 slots (5.98 GiB), 3004 of its 4608 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-29",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-29 (CUDA0), 30-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 93.2% (5367 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 93.7% (4049 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 96.7% (4178 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 97.7% (26272 hits / 26880 lookups)",
+  "strata serve: decode expert cache hit rate: 97.1% (4194 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (110748 hits / 113280 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (113787 hits / 114240 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (127856 hits / 128640 lookups)",
+  "strata serve: decode expert cache hit rate: 96.3% (116528 hits / 120960 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (116423 hits / 117120 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (120354 hits / 120960 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x3-auto.json

@@ -0,0 +1,89 @@
+{
+ "label": "v-x3-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,1,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9525.4,
+   "pp_tok_s": 1686.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15017.1,
+   "pp_tok_s": 1847.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4039.6,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 360.2,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 67.6,
+   "reps": [
+    59.8,
+    83.1,
+    59.9
+   ],
+   "hit_rate": 0.9967581085879247,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with the force of a battering ram. Elias stood in the lantern room, his hands rough as sandpaper, adjusting the massive Fresnel lens. The rotation mechanism groaned, a sound like a ship\u2019s anchor chain being h"
+  },
+  {
+   "name": "code",
+   "tok_s": 89.6,
+   "reps": [
+    81.0,
+    104.8,
+    83.0
+   ],
+   "hit_rate": 0.9975041387598007,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 2: 522 windows; per window: wait for the GPU 12.405 ms, pool + plan 0.242 ms, host staging 0.006 ms, commit 0.375 ms",
+  "strata serve: stage 0: 590 windows; per window: wait for the GPU 7.387 ms, pool + plan 0.549 ms, host staging 0.517 ms, commit 0.248 ms",
+  "strata serve: stage 1: 590 windows; per window: wait for the GPU 3.071 ms, pool + plan 0.011 ms, host staging 0.000 ms, commit 0.606 ms",
+  "strata serve: stage 2: 590 windows; per window: wait for the GPU 12.592 ms, pool + plan 0.250 ms, host staging 0.006 ms, commit 0.377 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session; 16.67 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.30 GiB for experts",
+  "strata generate: layer split auto: CUDA1 68 SMs at 1.54 GHz -> 0.69 ms per layer, 13.50 GiB for experts",
+  "strata generate: layer split auto: CUDA2 82 SMs at 1.86 GHz -> 0.48 ms per layer, 14.14 GiB for experts",
+  "strata generate: layer split auto: K=22,23 - predicted 19.9 ms per decode window; the caches hold 11298 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-22, expert cache 256 slots (0.52 GiB), 256 of its 256 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 23-47, expert cache 6400 slots (12.82 GiB), 6400 of its 6400 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-22 (CUDA1), 23-47 (CUDA2), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.7% (4256 hits / 4310 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (29862 hits / 30170 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4264 hits / 4313 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (134418 hits / 136339 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (116986 hits / 117525 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (129134 hits / 129554 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-auto-c1024.json

@@ -0,0 +1,84 @@
+{
+ "label": "x2-auto-c1024",
+ "extra": [
+  "--layer-split",
+  "auto",
+  "--prefill",
+  "1024"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2648.7,
+   "pp_tok_s": 1029.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5547.0,
+   "pp_tok_s": 1510.5,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10123.3,
+   "pp_tok_s": 1587.0,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 17652.0,
+   "pp_tok_s": 1571.7,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4379.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 235.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 82.7,
+   "hit_rate": 1.0,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+  },
+  {
+   "name": "code",
+   "tok_s": 111.1,
+   "hit_rate": 1.0,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: K=19 - the caches hold 12288 of 12288 profiled pairs (fullest device 99%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3613 slots",
+  "strata generate: expert cache 4864 slots, 8.76 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7424 slots (14.66 GiB), 7424 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 100.0% (5760 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (26400 hits / 26400 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (130080 hits / 130080 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (110400 hits / 110400 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-auto.json

@@ -0,0 +1,82 @@
+{
+ "label": "x2-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2711.7,
+   "pp_tok_s": 1005.3,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5010.3,
+   "pp_tok_s": 1672.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 8994.8,
+   "pp_tok_s": 1786.1,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15497.2,
+   "pp_tok_s": 1790.2,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4050.3,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 249.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.0,
+   "hit_rate": 0.9920220447608565,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. The massive Fresnel lens, a cathedral of glass and crystal, began its slow, hypnotic sweep"
+  },
+  {
+   "name": "code",
+   "tok_s": 101.8,
+   "hit_rate": 0.991767530342175,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: K=19 - the caches hold 11767 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4606 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7160 slots (14.14 GiB), 7160 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.5% (6141 hits / 6234 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (3797 hits / 3838 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (3804 hits / 3838 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3816 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (3825 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27693 hits / 27832 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3815 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (127081 hits / 128103 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (119386 hits / 120377 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-k16.json

@@ -0,0 +1,81 @@
+{
+ "label": "x2-k16",
+ "extra": [
+  "--layer-split",
+  "16"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2987.1,
+   "pp_tok_s": 912.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5908.6,
+   "pp_tok_s": 1418.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10655.8,
+   "pp_tok_s": 1507.7,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 18484.6,
+   "pp_tok_s": 1500.9,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4790.8,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 243.3,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 71.4,
+   "hit_rate": 0.9814261642156863,
+   "text": "The wind didn\u2019t just blow; it screamed. It clawed at the granite base of the tower, trying to pry the lighthouse from its bedrock. Inside, however, the air was still, smelling faintly of kerosene and old paper.\n\nElias had been the keeper of the Point for thirty-two years. He knew every creak in the floorboards, every draft that slipped through the window seals, and exactly how much oil the great Fresnel lens needed to turn its slow, hypnotic rotation.\n\nIt was 2:00 AM. The storm was at its peak. The sea was a chaotic black mirror, whipped into whitecaps that looked like jagged teeth.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His boots, heavy with salt and wear, thumped a steady rhythm against the stone. *Thump. Thump. Thump.*\n\nAt the top, the lantern room hummed with the vibration of"
+  },
+  {
+   "name": "code",
+   "tok_s": 86.4,
+   "hit_rate": 0.9798276439089693,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 16)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4096 slots, 7.17 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 16-47, expert cache 7124 slots (14.14 GiB), 7124 of its 8192 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-15",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-15 (CUDA0), 16-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5626 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (4237 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (4269 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4279 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (4298 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (28471 hits / 28800 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4282 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (128135 hits / 130560 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (117109 hits / 119520 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-k22.json

@@ -0,0 +1,81 @@
+{
+ "label": "x2-k22",
+ "extra": [
+  "--layer-split",
+  "22"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2537.7,
+   "pp_tok_s": 1074.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 4440.0,
+   "pp_tok_s": 1887.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7872.5,
+   "pp_tok_s": 2040.8,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13405.0,
+   "pp_tok_s": 2069.6,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3610.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 236.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 78.1,
+   "hit_rate": 0.9851122208995093,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and he had been the keeper of the Blackwood Point Light for forty of those years. The other thirty-three had been spent in a hospital bed, waiting for the end.\n\nEvery night, Elias climbed the one hundred and twelve steps of the spiral staircase. His knees, worn smooth by the friction of his own boots over decades, protested with a sharp ache. He counted them. *One, two, three\u2026* The rhythm was his heartbeat. If he lost the count, he lost the light.\n\nAt the top, the lantern room hummed with the low, electric thrum of the Fresnel lens. It was a cathedral of glass and brass, a machine that had never failed him. Not once. Not in four decades.\n\nHe polished the brass railing with a rag that smelled of lemon and"
+  },
+  {
+   "name": "code",
+   "tok_s": 104.3,
+   "hit_rate": 0.9842543225870746,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 22)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (4241 hits / 4306 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4262 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4290 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4319 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (28379 hits / 28721 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (4285 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (135713 hits / 137764 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (118518 hits / 120414 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2r-auto.json

@@ -0,0 +1,82 @@
+{
+ "label": "x2r-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "2,0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2804.2,
+   "pp_tok_s": 972.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5641.7,
+   "pp_tok_s": 1485.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10643.4,
+   "pp_tok_s": 1509.5,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 18363.5,
+   "pp_tok_s": 1510.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4517.2,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 245.6,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 78.9,
+   "hit_rate": 0.9913117283950618,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the fabric stiff with salt and ice. Outside the thick glass of the lantern room, the Atlantic was a churning cauldron of black and white, the waves crashing against the granite base with a violence that shook the very foundations of the tower. It was November, 1924, and the storm had been raging for three days.\n\nElias was not a man who feared the dark. He was a man who feared the silence.\n\nHe checked the clock on the wall. 11:42 PM. The rotation of the great Fresnel lens was steady, a rhythmic *whirrr-click, whirrr-click* that seemed to sync with his own heartbeat. He had been doing this for forty years. His hands were mapped with scars from rope burns and broken glass, his eyes permanently squinted against the glare of the bulb."
+  },
+  {
+   "name": "code",
+   "tok_s": 93.4,
+   "hit_rate": 0.990286044973545,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 57.7 GB/s -> pcie_frac 0.55",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 9.98 GiB free",
+  "strata generate: layer split auto: K=36 - the caches hold 11775 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 18.62 GiB free, 700 MiB reserved -> 6634 slots",
+  "strata generate: expert cache 8828 slots, 16.45 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 36-47, expert cache 2945 slots (5.99 GiB), 2945 of its 3072 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-35",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-35 (CUDA0), 36-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.6% (5678 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4283 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (4285 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (4306 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27700 hits / 27840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4295 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (128474 hits / 129600 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (119785 hits / 120960 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2t-auto.json

@@ -0,0 +1,99 @@
+{
+ "label": "x2t-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,1",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 4201.1,
+   "pp_tok_s": 648.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 9655.6,
+   "pp_tok_s": 867.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 17573.8,
+   "pp_tok_s": 914.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 30615.6,
+   "pp_tok_s": 906.2,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 7864.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 366.9,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 45.5,
+   "reps": [
+    45.5
+   ],
+   "hit_rate": 0.9703345151405784,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, knew every bolt, every lever, and every creak in the iron bones of the tower.\n\nFor forty years, Elias had been the guardian of the Point. Not just of the light, but of the silence between the waves.\n\nIt was November. The sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that shook the floorboards beneath Elias\u2019s boots.\n\nHe climbed the stairs. One hundred and twelve steps. He counted them by feel, his left hand trailing along the cold iron railing. At the top, the air was"
+  },
+  {
+   "name": "code",
+   "tok_s": 58.0,
+   "reps": [
+    58.0
+   ],
+   "hit_rate": 0.9700542735219787,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 168 windows; per window: wait for the GPU 9.683 ms, pool + plan 0.509 ms, host staging 0.714 ms, commit 0.320 ms",
+  "strata serve: stage 1: 168 windows; per window: wait for the GPU 25.253 ms, pool + plan 0.881 ms, host staging 0.000 ms, commit 0.559 ms",
+  "strata serve: stage 0: 242 windows; per window: wait for the GPU 9.860 ms, pool + plan 0.550 ms, host staging 0.750 ms, commit 0.328 ms",
+  "strata serve: stage 1: 242 windows; per window: wait for the GPU 26.029 ms, pool + plan 0.980 ms, host staging 0.000 ms, commit 0.571 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=20 - the caches hold 10761 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 6157 slots (12.19 GiB), 6157 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=20 - the caches hold 10761 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 6157 slots (12.19 GiB), 6157 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 96.6% (5559 hits / 5754 lookups)",
+  "strata serve: decode expert cache hit rate: 97.5% (4205 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (4217 hits / 4312 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (4257 hits / 4318 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x3-auto.json

@@ -0,0 +1,99 @@
+{
+ "label": "x3-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2,1",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 3657.4,
+   "pp_tok_s": 745.3,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 7594.1,
+   "pp_tok_s": 1103.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 13996.4,
+   "pp_tok_s": 1147.9,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 24255.2,
+   "pp_tok_s": 1143.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 6104.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 393.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 51.9,
+   "reps": [
+    51.9
+   ],
+   "hit_rate": 1.0,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+  },
+  {
+   "name": "code",
+   "tok_s": 69.7,
+   "reps": [
+    69.7
+   ],
+   "hit_rate": 1.0,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 2: 165 windows; per window: wait for the GPU 13.452 ms, pool + plan 0.162 ms, host staging 0.000 ms, commit 0.328 ms",
+  "strata serve: stage 0: 234 windows; per window: wait for the GPU 6.235 ms, pool + plan 0.145 ms, host staging 0.704 ms, commit 0.214 ms",
+  "strata serve: stage 1: 234 windows; per window: wait for the GPU 15.592 ms, pool + plan 0.246 ms, host staging 0.000 ms, commit 0.657 ms",
+  "strata serve: stage 2: 234 windows; per window: wait for the GPU 13.916 ms, pool + plan 0.177 ms, host staging 0.000 ms, commit 0.335 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session; 18.62 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=12,33 - the caches hold 12288 of 12288 profiled pairs (fullest device 66%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3072 slots",
+  "strata generate: expert cache 3072 slots, 5.41 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 12-32, expert cache 5376 slots (10.22 GiB), 5376 of its 5376 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 33-47, expert cache 3840 slots (7.79 GiB), 3840 of its 3840 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-11",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-11 (CUDA0), 12-32 (CUDA1), 33-47 (CUDA2), one hand-off per window",
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session; 18.62 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=12,33 - the caches hold 12288 of 12288 profiled pairs (fullest device 66%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3072 slots",
+  "strata generate: expert cache 3072 slots, 5.41 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 12-32, expert cache 5376 slots (10.22 GiB), 5376 of its 5376 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 33-47, expert cache 3840 slots (7.79 GiB), 3840 of its 3840 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-11"
+ ]
+}

data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.json

@@ -0,0 +1,26 @@
+{
+ "mode": "project",
+ "dir": "per-layer",
+ "scale": 1.0,
+ "layer_range": [
+  4,
+  44
+ ],
+ "refusal_harmful_thinkoff": 2.0,
+ "refusal_harmful_thinkon": 2.0,
+ "refusal_harmless_thinkoff": 4.0,
+ "refusal_harmless_thinkon": 6.0,
+ "broken": 0,
+ "kl_mean": 0.18598,
+ "kl_max": 18.26549,
+ "kl_p99": 3.143571,
+ "decode_tps": 51.19,
+ "vector_file": "Qwen3.8-Flash-Next-experimental-speed-projection.gguf",
+ "base_model_gguf": "unsloth/Qwen3.8-Flash-Next-GGUF UD-Q2_K_XL",
+ "base_model_sha256_shard1": "a4f3b21e77353999829f2f767e9ac21ce9c71d29a74f2cc9eda48c9bf23c8b86",
+ "worktree_commit": "37e026e3b",
+ "criterion_status": "PASS",
+ "dirty_direction": false,
+ "note": "projection ablation h -= s*(h.v)*v, v unit-normalized per layer. dir=per-layer",
+ "branch": "cvec-hooks (github.com/Cudila/llama.cpp)"
+}

data/experimental-speed-projection/README.md

@@ -0,0 +1,21 @@
+# experimental-speed-projection (EXPERIMENTAL, off by default)
+
+`Qwen3.8-Flash-Next-experimental-speed-projection.gguf` is a 480 KB control vector for Qwen3.8-Flash-Next: one unit
+direction of 2,560 values per layer (`direction.1` .. `direction.47`, GGUF `controlvector` format). The JSON beside it
+is the configuration its authors selected and their scores.
+
+**What it does.** Turned on, Strata removes each direction from the residual stream after layers 4-44
+(`h -= (h . v) v` on every hyper-connection stream), the same operation as llama.cpp's `--cvec-mode project` in the
+package's patches. The package's own documentation describes the vector as a **refusal-direction projection**: the
+model declines far fewer requests (it reports 1 of 50 vs 50 of 50 on its test set), and removing refusals removes a
+safety behaviour - you are responsible for what the model writes with it on. It also shifts the model's predictions
+slightly on ordinary text (measured here: `bench/results/2026-09-27-esp/`). It is not an optimization in the engine:
+on the same text it costs 0.2-0.4% per token; a chat's speed with it on depends on the text the model writes.
+
+**Turning it on.** `START-HERE.bat --setup` asks (default: off), or pass `--experimental-speed-projection on`. With it
+loaded, the web app's Sampling drawer and the API field `"experimental_speed_projection": false` switch it off per
+request. See `docs/DETAILS.md`, "Experimental speed projection".
+
+**Origin and license.** A local package of the vector (the published original uses filenames containing "refusal";
+the bytes are the same), made from Qwen3.8-Flash-Next activations, so the Qwen Community License 1.0 of the model
+applies to it. Made for the original Qwen3.8-Flash-Next (not Swift 1.5).

docs/DETAILS.md

@@ -4,42 +4,81 @@ The technical side of Strata: every measured number, the API, images, all settin
 New here? Start with the [README](../README.md) - it has everything you need to install and use it.
 
 > **On this page:** [Speed](#speed-measured) · [Other GPUs](#other-gpus-estimated) · [Which model?](#which-model) ·
-> [Requirements](#before-you-start) · [Windows](#windows) · [Linux](#linux) · [API](#using-it) · [Images](#images-vision) ·
+> [Requirements](#before-you-start) · [Windows](#windows) · [Linux](#linux) · [API](#using-it) ·
+> [MCP tools](#tools-from-mcp-servers) · [Images](#images-vision) ·
 > [Troubleshooting](#troubleshooting) · [How it works](#how-it-works)
 
 ---
 
 ## Speed (measured)
 
-RTX 5070 **12 GB**, Ryzen 5 7600 (6 cores), 64 GB DDR5-5200, Windows. One code-agent prompt per length, 256 generated
-tokens, MTP speculative decoding on. "262K" is the model's full context window (a 259,943-token prompt).
+RTX 5070 **12 GB**, Ryzen 5 7600 (6 cores), 64 GB DDR5-5200, Windows, engine 0.1.14 with the settings setup writes
+(`--prefill auto`, 8-bit KV above 4K, KV streaming from 64K). One code-agent prompt per length, 256 generated tokens,
+MTP speculative decoding on. "262K" is the model's full context window (a 259,943-token prompt). The IQ2_XS row was
+measured with Swift 1.5's IQ2_XS, which runs at the original's speed.
 
 ### Prompt processing (tokens/s)
 
 | Model | 1K | 4K | 32K | 64K | 128K | 262K |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: |
-| **Q2_0** | 389 | 539 | 571 | 561 | 543 | 496 |
-| **IQ2_XS** | 332 | 463 | 495 | 486 | 472 | 437 |
-| **IQ3_XXS** | 285 | 410 | 435 | 427 | 414 | - |
+| **Q2_0** | 494 | 1,007 | 1,308 | 1,294 | 1,208 | 967 |
+| **IQ2_XS** | 461 | 811 | 1,238 | 1,136 | 1,071 | 886 |
+| **IQ3_XXS** | 415 | 770 | 1,108 | 1,065 | 1,015 | - |
+| **IQ3_S** | 396 | 737 | 1,070 | 1,070 | 931 | - |
+| **Coder** | 599 | 1,152 | 1,298 | 1,350 | 1,266 | 1,034 |
 
 ### Output (tokens/s)
 
 | Model | 1K | 4K | 32K | 64K | 128K | 262K |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: |
-| **Q2_0** | 88.7 | 94.6 | 87.5 | 76.4 | 65.1 | 56.3 |
-| **IQ2_XS** | 82.0 | 78.0 | 65.3 | 63.7 | 52.0 | 48.0 |
-| **IQ3_XXS** | 64.6 | 65.6 | 57.3 | 54.4 | 44.8 | - |
+| **Q2_0** | 84.3 | 90.3 | 73.6 | 69.0 | 67.2 | 60.3 |
+| **IQ2_XS** | 74.4 | 73.8 | 71.5 | 64.3 | 59.8 | 52.8 |
+| **IQ3_XXS** | 60.3 | 62.1 | 51.4 | 50.0 | 45.8 | - |
+| **IQ3_S** | 51.5 | 51.6 | 48.2 | 48.8 | 40.5 | - |
+| **Coder** | 53.3 | 50.6 | 53.3 | 50.8 | 44.0 | 42.8 |
 
-IQ3_XXS at 262K is not measured: with its 43 GB of experts, a 260K-token context brings a 64 GB PC to its memory
-limit. Use up to 128K with IQ3_XXS on 64 GB.
+Output speed depends on the text as well: speculative decoding runs faster when more of the drafted tokens are
+accepted, so a different answer to the same prompt moves it by several percent. Run back to back on the 4K prompt,
+0.1.14 writes 88.5 tokens/s and 0.1.12 85.7. The numbers before 0.1.13 (prompts about half as fast):
+[`bench/results/2026-09-24-final`](../bench/results/2026-09-24-final/matrix.md); these:
+[`bench/results/2026-09-28-speed-0114`](../bench/results/2026-09-28-speed-0114/README.md).
 
-Time to first token is prompt length / prompt speed: about 7 s at 4K, 55 s at 32K, 4 minutes at 128K and 9 minutes at
-262K. The raw numbers: [`bench/results/`](../bench/results/). The [paper](paper/Strata-Paper.pdf) explains every number.
+IQ3_XXS and IQ3_S at 262K are not measured: with their 43 / 50 GB of experts, a 260K-token context brings a 64 GB PC
+to its memory limit. Use up to 128K with them on 64 GB (setup caps it). IQ3_S (engine 0.1.4 or newer) is only published
+for the original model, not for Swift 1.5.
+
+**KV streaming (engine 0.1.5):** at 64K and more, setup keeps the context's KV cache in RAM and only the part the
+attention reads in VRAM (`--kv-resident 32768`), so more experts fit on the GPU. Q2_0 at 262K: 50.9 -> 62.6 tokens/s
+(1,589 -> 3,872 experts in VRAM); at 128K about +6%. The attention reads exactly the same values (only where the KV lives changes); it
+costs ~13.7 KB of RAM per context token (1.7 GB at 128K). Existing installs: run `START-HERE.bat --setup` once to turn
+it on.
+
+**4-bit KV cache (engine 0.1.8, optional):** `START-HERE.bat --setup` asks above 8K context (or pass `--kv q4_0`). It
+halves the KV cache's memory with a Hadamard rotation before 4-bit rounding (PR #21), about 4% faster at 128K, but it
+is measurably less precise on long documents (perplexity +8-12%; needle tests still pass). 8-bit stays the default.
+Details: [`bench/results/2026-09-27-kv-q4`](../bench/results/2026-09-27-kv-q4/README.md).
+
+Time to first token is prompt length / prompt speed: with Q2_0 about 4 s at 4K, 25 s at 32K, under 2 minutes at 128K
+and 4.5 minutes at 262K (engine 0.1.13 made long prompts about twice as fast, below).
+
+**Faster prompts (engine 0.1.13):** the prompt is read in chunks of up to 8,192 tokens instead of 2,048 (`--prefill
+auto`: the largest chunk whose buffers fit in the expert-cache slots it borrows, and a request borrows only what its
+prompt needs); the experts are multiplied by llama.cpp's quantized MMQ kernels instead of being expanded to FP16
+first; the next layer's experts stream over PCIe while the current layer's attention runs; the PLE block runs for the
+whole chunk at once; unpinned experts are copied by helper threads. Measured on the RTX 5070 12 GB, 64 GB RAM,
+32K-token prompt: Q2_0 572 -> 1,290 tokens/s, IQ3_S 383 -> 1,208. Through the server (Q2_0, 128K context): 999 tokens
+353 -> 438 tokens/s, 6,927 tokens 529 -> 1,077, 28,584 tokens 584 -> 1,249. Output speed is unchanged. Needles 5/5
+(1K-262K). Details and the quality check:
+[`bench/results/2026-09-28-prefill-speed`](../bench/results/2026-09-28-prefill-speed/README.md). Existing installs
+switch to `--prefill auto` the next time START-HERE / setup.sh starts them. The raw numbers:
+[`bench/results/`](../bench/results/). The [paper](paper/Strata-Paper.pdf) explains every number.
 
 ## Other GPUs (estimated)
 
 Not measured - estimated from the runs above (same CPU and 64 GB RAM): the GPU part scaled by memory bandwidth, the CPU
 part by how many more experts the card's VRAM holds. Treat as **±20%**. Numbers are *prompt / output* tokens/s.
+The prompt figures predate engine 0.1.13, which about doubled prompt speed on the measured card; how much of that a
+card gains depends on its PCIe link (the experts stream over it), so they are still the older estimates.
 
 | GPU | Model | 1K | 4K | 32K | 64K | 128K | 262K |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
@@ -51,7 +90,9 @@ part by how many more experts the card's VRAM holds. Treat as **±20%**. Numbers
 |  | IQ3_XXS | ~260 / ~106 | ~374 / ~103 | ~396 / ~89 | ~390 / ~85 | ~378 / ~71 | - |
 
 More VRAM matters more than a faster GPU: every extra GB holds ~700 more experts, and every expert on the GPU is one the
-CPU does not have to compute. A 3090's 24 GB takes most of the CPU work away.
+CPU does not have to compute. A 3090's 24 GB takes most of the CPU work away. (Since 0.1.14 the expert profile ranks
+all 24,576 experts; before, the cache stopped at 8,000, about 10-14 GB. `tools/make_profile.py` builds a profile from
+your own prompts: run the engine once with `--dump-routing trace.bin`, see the tool's help.)
 
 ## Which model?
 
@@ -64,7 +105,24 @@ of [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).
 | **IQ2_XS** | 68 GB | ~36 GB experts + ~6 GB | close to Q2_0 | a bit better |
 | **IQ3_XXS** | 76 GB | ~43 GB experts + ~6 GB | slower (more CPU work) | best |
 
-With 64 GB of RAM all three fit (close the browser for IQ3_XXS, and keep its context at 128K or less). With 48 GB only Q2_0 / IQ2_XS may fit. 32 GB is not enough.
+With 64 GB of RAM all three fit (close the browser for IQ3_XXS, and keep its context at 128K or less). With 48 GB only Q2_0 / IQ2_XS may fit. With 32 GB: the Coder (below).
+
+### Or: the Coder (half the experts, for code)
+
+**[Qwen3.8-Flash-Next GSQ-RCO Coder](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF)** is
+ISTA-DASLab's expert-pruned release: 256 of each layer's 512 experts are kept (still 10 active per token), chosen with
+RCO on code, agentic and vision calibration data; its authors report 91.3% of the full model's SWE-bench Verified and
+98.7% of LiveCodeBench v6. One size, named IQ1_M for its 1.89 bits per *original* parameter; the kept experts are
+stored like IQ3_S (IQ2_S-IQ4_XS gate/up, IQ4_NL/Q2_0 down). Shard 1 is 29.6 GB (experts: 23 GB of RAM), so it runs
+on **32 GB of RAM**, and at 262K on 64 GB. Its shard 2 and its vision encoder are the original's files: with the
+original installed, setup downloads only shard 1. Strata ships its expert profile (`data/expert-profile-coder.bin`,
+the shipped ranking mapped onto the kept experts through the release's `rco-allocation.txt`: 72% of the expert
+reads hit the GPU on a 12 GB card). Images work; the experimental speed projection loads and runs on it (it was made
+for the full model).
+
+```
+START-HERE.bat --setup --family coder
+```
 
 ### Or: Swift 1.5 (a fine-tune that thinks shorter)
 
@@ -96,7 +154,11 @@ You need **only an NVIDIA driver** (version 580 or newer; update it with the NVI
 | Disk | ~70-80 GB free for the model, ~6 GB for the MTP layer (+1 GB with images). **Q2_0 on an AVX-512 CPU** also writes a one-time ~40 GB copy of its experts for the fast CPU kernel. An NVMe SSD is strongly recommended. |
 | OS | Windows 10/11, or Linux (Ubuntu 22.04/24.04 get everything installed automatically). |
 
-What the first start installs, all inside this folder (`.venv/`, `engine/`, `third_party/`, `models/`, `packs/`, `mtp/`):
+What the first start installs: in this folder `.venv/`, `engine/` and `third_party/`; the model files (`models/`,
+`packs/`, `mtp/`, 70-120 GB) in **`Strata-data` next to this folder**, so a new copy of Strata (an update unzipped
+elsewhere) finds them and sets itself up the same way. The place is remembered per user (`%APPDATA%\Strata\settings.json`,
+`~/.config/strata/settings.json`); `--data-dir` chooses another. Installs from before 0.1.16 are moved there by the next
+start (a rename on the same drive; files on another drive are used where they are).
 Python 3.12 if you have none (for your user account, no admin), a private Python environment, NVIDIA's CUDA libraries
 (from pip, ~0.4 GB), the ready-made Strata engine for RTX 30/40/50, the model and the MTP draft layer. If no
 ready-made engine fits your PC, it offers to install the build tools (Visual Studio Build Tools + CUDA Toolkit on
@@ -116,21 +178,44 @@ Windows, `build-essential` + CUDA on Ubuntu) and compiles the engine for your GP
 4. **Images?** yes / no (see [Images](#images-vision)).
 
 Then it downloads and prepares everything (the model is 66-76 GB, so the first start takes a while; an interrupted
-download continues where it stopped) and **starts the model**: your browser opens `http://127.0.0.1:8080`, a small
-page that shows it is running and lets you chat. The API is at `http://127.0.0.1:8080/v1` for your apps.
+download continues where it stopped) and **starts the model**: your browser opens `http://127.0.0.1:8080`, the Strata
+app. It has three tabs:
+- **Chat:** streaming answers, the model's thinking (folded away once it answers), code with a copy button, pictures when
+  images are on, and sampling and thinking-level settings. Chats stay in your browser.
+- **Monitor:** what the model is doing (reading the prompt, with progress, or writing, at how many tokens/s); GPU load,
+  VRAM, temperature, power and PCIe traffic; CPU, RAM and disk; the context in use; the last requests.
+- **About:** the model and engine settings, and the addresses to connect other apps.
+
+`http://127.0.0.1:8080/?q=your question` opens it with a new chat already asking. The API is at
+`http://127.0.0.1:8080/v1` for your apps.
 
 **Every time after that**, `START-HERE.bat` just starts the model (30-90 s to load 34-43 GB into RAM). Nothing is
 downloaded again. Closing the window stops the model.
 
 ```
 START-HERE.bat --setup                          install another model, or change context / images
+SETUP.bat                                       the same (double-click it)
 START-HERE.bat --model IQ2_XS --context 32768 --vision yes --yes     no questions
 START-HERE.bat --gguf-dir D:\models\IQ2_XS       use GGUF files you already have
+START-HERE.bat --data-dir E:\Strata-data         keep the model files somewhere else
 START-HERE.bat --port 8081                      another port
+START-HERE.bat --gpu 1                          another GPU (numbered as nvidia-smi; setup picks the one with the most VRAM)
+START-HERE.bat --calibrate                      tune the engine for this PC (about 5-10 minutes), then start
 ```
 
 With more than one model installed, it asks which one to start. `run-<model>.bat` starts a model directly.
 
+**Tuning for your PC (`--calibrate`, engine 0.1.19).** Three engine settings depend on the PC more than on the model:
+- the share of the experts missing from VRAM that are copied to the GPU instead of computed by the CPU
+  (`--pcie-frac`: a fast PCIe link and a slower CPU want more, a laptop's narrower link less);
+- how sure the draft layer must be to add another guess to a check (`--spec-min-p`);
+- how many CPU threads compute experts (`--pool-workers`: on CPUs with efficiency cores, fewer can be faster).
+
+The defaults were measured on a Ryzen 5 7600 with an RTX 5070. Setup offers to measure them on your PC after an
+install; `START-HERE.bat --calibrate` (Linux: `./setup.sh --calibrate`) does it any time. It measures the output
+speed with each setting and keeps one only when it is more than 3% faster. The result is remembered per PC and model
+(in the settings file next to the data folder's record), so updates keep it.
+
 ### Chat in the terminal (optional)
 
 ```
@@ -150,6 +235,15 @@ for the build tools), and the same start: `http://127.0.0.1:8080`. Later runs of
 start the model directly. Options as on Windows (`./setup.sh --setup`, `--model Q2_0 --yes`, `--gguf-dir /data/Q2_0`).
 Terminal chat: `.venv/bin/python chat.py`.
 
+- **Updating:** `git pull`, then `./setup.sh`: it compiles the engine again when its source changed (a minute or
+  two for the changed files). If that compile fails, it says so and starts the engine you had.
+- **Other distributions** (Arch, Fedora, ...): install the C++ compiler and the CUDA Toolkit 13 with your package
+  manager first (Arch: `sudo pacman -S base-devel cuda`); setup finds `nvcc` on PATH, in `/usr/local/cuda*` and in
+  `/opt/cuda*`, and does the rest.
+- **WSL** works (Ubuntu 24.04 tested), with one limit: the NVIDIA driver pins only about 1 GB of RAM there, so KV
+  streaming (`--kv-resident`) is off and the KV cache stays in VRAM, and the experts are copied to the GPU from
+  unpinned RAM (slower prompts than native Linux).
+
 ---
 
 ## Using it
@@ -162,6 +256,8 @@ The server listens on `http://127.0.0.1:8080` (change with `--port` in setup, or
 | Anthropic Messages (stream and non-stream, tools) | `POST /v1/messages` |
 | Model list / health | `GET /v1/models`, `GET /health` |
 | What the model is doing right now | `GET /status` |
+| Everything the Monitor tab shows (engine, live state, last requests, hardware) | `GET /metrics` |
+| The MCP servers, their state and tools ([below](#tools-from-mcp-servers)) | `GET /mcp` |
 
 ```bash
 curl http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
@@ -193,16 +289,91 @@ print(r.choices[0].message.content)
   15 s, and `GET /status` says what it is doing (`reading the prompt`, `answering`, tokens so far). Closing the
   connection or pressing stop in your app really stops the model, so the next request starts at once.
 - **Chat apps.** Any app with an "OpenAI-compatible" provider works: base URL `http://127.0.0.1:8080/v1`, any API key.
-- **Context.** Chosen in setup (8K-262K). Requests longer than that are refused, never silently cut.
-- **From other devices / the internet.** The server listens on your PC only (`127.0.0.1`). To reach it from elsewhere,
-  put a tunnel in front of it, for example [cloudflared](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/):
+- **Claude Code** (Strata 0.1.17 or newer): set `ANTHROPIC_BASE_URL=http://127.0.0.1:8080` and
+  `ANTHROPIC_MODEL` to a Claude model name it knows (it refuses names it doesn't; Strata ignores the name), plus any
+  `ANTHROPIC_AUTH_TOKEN` (or your `api_key`, if you set one).
+- **Context.** Chosen in setup (8K-262K). Requests longer than that are refused, never silently cut. A request whose
+  `max_tokens` would run past the context is refused too (400); agents that always ask for their full output cap
+  can instead get it shortened to the room left: add `"fit_max_tokens": true` to `strata-<model>.json` (or pass
+  `--fit-max-tokens` to `serve/server.py`). A prompt that leaves no room at all is still refused.
+- **From other devices on your network.** The server listens on your PC only (`127.0.0.1`) unless you say otherwise:
+  run setup with `START-HERE.bat --setup --host 0.0.0.0 --api-key some-long-secret` (or add `"host": "0.0.0.0"` and
+  `"api_key": "..."` to `strata-<model>.json`). The server window then prints this PC's addresses
+  (`from other devices: http://192.168.x.x:8080/`); open that on the other device, or use `.../v1` as an API base URL.
+  On Windows the firewall blocks it until you allow it: accept its prompt for Python (private networks), or run
+  `New-NetFirewallRule -DisplayName "Strata 8080" -Direction Inbound -Protocol TCP -LocalPort 8080 -Action Allow -Profile Private`
+  in an admin PowerShell, and make sure the network is set to Private.
+- **From the internet.** Put a tunnel in front of it, for example [cloudflared](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/):
   `cloudflared tunnel --url http://127.0.0.1:8080`. **Set a key first**, or anyone with the link can use your PC:
   add `"api_key": "some-long-secret"` to `strata-<model>.json` (or set the `STRATA_API_KEY` environment variable);
   clients then send it as their API key.
 
-**Current limits (v1):** one request at a time; greedy decoding (temperature is ignored); every request processes its
-whole prompt again (no conversation cache yet, so long chats have a long time-to-first-token); images only when set up
-with them (below); no video.
+**Conversation cache.** A request that continues a chat reads only the part after what the engine already holds: the
+live session, or one of the checkpoints it keeps in RAM (up to 6, ~118 MB each, taken at the start of each new
+assistant turn and every 16K prompt tokens). A checkpoint is used only when the prompt starts with exactly its tokens
+and pictures. The oldest checkpoint - in practice the end of the system prompt, which every chat of the same client
+shares - is kept for good while the rest rotates by least recent use, so a NEW chat that shares that prefix starts
+reading after it instead of from token 0. A prompt read from the start is also checkpointed at the end of its system
+prompt when that is 2,048 tokens or more (engine 0.1.20; PR #62 + #65), so that root exists for agent clients with long
+system prompts and tool lists. Engine options: `--prompt-cache N` (0 = off), `--prompt-cache-every N`,
+`--prompt-cache-root N` (0 = no system-prompt checkpoint), `--turn-token ID`.
+
+**Current limits (v1):** one request at a time, and one conversation's history in the KV cache at a time (switching
+between two chats re-reads the part where they diverge; the shared prefix, such as the system prompt, is reused); images
+only when set up with them (below); no video. **Temperature / top_p / top_k / min_p /
+seed** are honored per request (OpenAI and Anthropic fields); with the default adaptive expert tier a sampled result
+is not reproducible run to run - for seed-reproducible output add `--adapt-every 100000` (static residency) to the
+engine arguments. The run config's optional `sampling` block sets the defaults for requests that leave the fields out
+(`"sampling": {"temperature": 1.0, "top_p": 0.95, "top_k": 20}`); a request's own fields always win, and with no
+block at all a request without sampling keys decodes greedy. The penalties (`presence_penalty`, `frequency_penalty`,
+`repetition_penalty`, with `penalty_last_n` capping how many recent tokens they count over, default 64 when any
+penalty is set) ride the same path; they count the tokens the request has consumed, so a repetition penalty
+suppresses what the model itself just said, not the prompt alone. Since engine 0.1.19 they apply to every token
+the speculative decoding checks at once, exactly as if it decoded one token at a time (before, only the first of
+each batch got them). That makes requests with penalties 1-11% slower than in 0.1.18: the draft layer guesses
+without penalties, so more of its guesses are now rejected. Requests without penalties are unchanged. `top_k` keeps at most 64 candidates: `0` ("off") or anything above 64 uses all 64.
+
+---
+
+## Tools from MCP servers
+
+The chat page can give the model tools from [MCP](https://modelcontextprotocol.io) servers, as LM Studio and Claude
+Desktop do: reading your files, fetching web pages, searching, anything an MCP server offers. List the servers in
+`strata-<model>.json` under `"mcp_servers"` - the same shape as Claude Desktop's `mcpServers` block, which you can
+also paste as it is (key `"mcpServers"`):
+
+```json
+"mcp_servers": {
+  "files": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "C:\\Users\\me\\Documents\\notes"]},
+  "search": {"url": "http://127.0.0.1:3000/mcp", "headers": {"Authorization": "Bearer ..."}}
+},
+"mcp": {"timeout_s": 60, "max_result_chars": 20000, "max_rounds": 8}
+```
+
+Or keep them in their own file and start the server with `--mcp-config path\to\claude_desktop_config.json` (a file
+with an `mcpServers` block; add it to the `serve/server.py` line of your run script). Restart Strata after a change.
+
+- **A program** (`command`, `args`, optional `env` and `cwd`) is started by Strata and spoken to over its
+  stdin/stdout; `npx`, `uvx`, `python` and friends are found on `PATH` as usual (Node.js is needed for `npx`
+  servers). **An address** (`url`, optional `headers`) uses MCP's Streamable HTTP transport (the older SSE-only
+  transport is not supported). `"disabled": true` leaves an entry out.
+- The servers start with Strata, in the background; the server window says what each one offers
+  (`MCP server 'files': 14 tools (...)`), or why it did not start - its tools are then left out and the chat works
+  without them. The Monitor tab lists them, and the Sampling drawer has **Use tools from MCP servers** (on by
+  default). A server that stops later is started again at its next call.
+- In the chat each call shows as a small block (tool, arguments, result); the model reads the result and goes on,
+  up to `max_rounds` calls in a row per answer. A tool that fails or takes longer than `timeout_s` (default 60 s)
+  gives the model an `error: ...` result instead of ending the chat. Results longer than `max_result_chars`
+  (default 20,000 characters) are cut, with a note, before the model reads them. Stop stops a running tool too.
+- Only the chat page uses them. API clients (omp, Claude Code, OpenAI and Anthropic SDKs) see the API exactly as
+  before and keep their own tools; a request to `/v1/chat/completions` opts in with `"strata_mcp": true` (it then
+  gets `strata_mcp` tool events in the stream).
+
+**Security.** MCP tools run on your PC with your user's rights, and **the model decides when to call them** - also
+because of what it reads (a web page or a file can contain instructions). Give a filesystem server only the folders
+it needs, prefer read-only tools, and don't add servers you don't trust. The tools can only be used from the chat
+page itself (a request with another site's Origin or without a JSON content type is refused); if Strata is reachable
+from other devices, set an API key.
 
 ---
 
@@ -280,6 +451,45 @@ test images.
 
 ---
 
+## Experimental speed projection (EXPERIMENTAL, off by default)
+
+**This is an experiment, not a finished feature.** It ships with Strata but stays off unless you turn it on.
+
+A 480 KB control vector for Qwen3.8-Flash-Next (`data/experimental-speed-projection/`, see its README). After each
+of layers 4-44 the engine removes one direction from every hyper-connection stream of the residual: `h -= (h . v) v`,
+one unit vector `v` per layer, exactly as llama.cpp does with the package's `--cvec-mode project` patches.
+
+**What it changes.** The vector's own package describes it as a **refusal-direction projection**: with it the model
+declines far fewer requests (it reports 1 of 50 vs 50 of 50 on its test set), and removing refusals removes a safety
+behaviour - you are responsible for what the model writes with it on. It also shifts ordinary answers a little
+(measured below). It is not an optimization in the engine: on the same text it costs 0.2-0.4% per token. What a
+chat's tokens/s does with it on depends on the text the model writes (length, repetition, how well the drafts land),
+so measure it on your own prompts; the Monitor marks every request ESP or stock.
+
+**Turning it on (at setup).** `START-HERE.bat --setup` asks "Turn on the experimental speed projection?" (default:
+no), or pass `--experimental-speed-projection on` (`off`, or a path to another vector GGUF). Only for the original
+Qwen3.8-Flash-Next, not Swift 1.5. It writes these engine flags (llama.cpp's) into `strata-<model>.json`:
+
+```
+--control-vector-scaled <Strata>\data\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0
+--control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer
+```
+
+The engine log then says `control vector mode = project, dir = per-layer, layers 4..44 (41 steered)`, and the web
+app's About tab lists it. (`--cvec-mode add` is llama.cpp's stock additive mode, for additive vectors.)
+
+**Per request.** A loaded vector is on for every request unless it says otherwise: the web app's Sampling drawer has
+a switch, and the API takes `"experimental_speed_projection": false` in the request body (OpenAI and Anthropic; a
+config default goes in `"sampling": {"experimental_speed_projection": false}`). Switching drops the conversation
+cache once, since the model state was computed the other way. Switched off, the output is token-for-token the stock
+model's.
+
+**Measured here** (Q2_0, fixed experts, 2,557 teacher-forced tokens of code, a document and a chat): the top-1 token
+changes at 10% of positions, mean KL from the stock model 0.063 nats (max 4.1), perplexity +15% on code, +2.3% on
+the document, +0.4% on the chat. Details: `bench/results/2026-09-27-esp/`.
+
+---
+
 ## Troubleshooting
 
 | Symptom | What to do |
@@ -288,14 +498,19 @@ test images.
 | Python or the build tools could not be installed | Install what it names (links are printed), then run it again. Everything already done is kept. |
 | `port 8080 is already in use` | Strata is already running (look for its window), or another program uses the port: `START-HERE.bat --port 8081`. |
 | `cudaHostRegister ... out of memory` in the log | Normal on Windows: the engine pins the experts in per-layer slices instead. Only a problem if the load then fails. |
-| The first start takes minutes | It is reading 34-43 GB into RAM; the second start is faster while the files are in the OS cache. |
+| `ExpertCache: cudaMalloc(...) failed: out of memory` although VRAM is free | Windows' page file is off or tiny: every allocation on the graphics card is also charged to Windows' commit (RAM + page file). Set the page file to "System managed" (System > About > Advanced system settings > Performance > Advanced > Virtual memory) and restart. Since 0.1.19 the engine retries with a smaller cache instead of stopping, and setup warns about a page file under 4 GB (issue #60). |
+| The first start takes minutes | It is reading 34-55 GB into RAM; the second start is faster while the files are in the OS cache. |
+| The PC freezes for a few minutes at the start | Normal, most of all the first time (the server window says when it happens): the engine loads the experts into RAM, pins part of it for the GPU and sizes the expert cache. Wait; don't close the window. Still frozen after 10 minutes: restart the PC, close other programs, try again, or pick a smaller size. |
+| `the engine stopped unexpectedly (exit code ...)` | The engine process ended mid-answer - usually out of RAM (Linux ends the biggest program: `sudo dmesg \| grep -i -E 'killed process\|out of memory'`). The next request starts it again by itself. If it repeats: close other programs or pick a smaller size. The server also warns at start when the model's experts leave less than ~6 GB of RAM for everything else. |
 | Slow output, disk light busy | Not enough free RAM: close other programs, or choose Q2_0 / IQ2_XS. |
 | `prompt ... exceeds the context` | The request is longer than the context you chose: run setup again with a bigger `--context`. |
 | Slower than the tables | The monitor plugged into the GPU and other GPU programs take VRAM from the expert cache; RAM running below its rated speed (enable EXPO/XMP in the BIOS) slows the CPU half. |
 | `this server was started without the vision encoder` | The model was set up for text only: run setup again with `--vision gpu`. |
 | A picture is refused or `cannot read the image` | The file is not a picture Pillow can open (JPEG, PNG, WebP, GIF, BMP, TIFF, AVIF work). |
 | Pictures are slow (10-30 s) | The encoder runs on the CPU: run setup again with `--vision gpu` (needs ~1.4 GB of VRAM). |
-| A request never finishes (older engines, mostly with pictures) | Run `START-HERE.bat` once to get engine 0.1.2 or newer. It keeps a real margin of VRAM free: the log says `... MiB of VRAM free with everything loaded` and warns when it is close to 0. |
+| A request never finishes: "reading the prompt", GPU "100%" at low power | The GPU ran out of VRAM (engines before 0.1.9 could end with ~30 MiB free at large contexts). Run `START-HERE.bat` once to get engine 0.1.9 or newer; the log then says `... MiB of VRAM free with everything loaded` (a few hundred) and names the `--vram-reserve-mib` to add if it is low. |
+| Generation stops mid-answer, GPU "100%", one CPU core busy | Fixed in engine 0.1.12 (issue #29, a race in the CPU expert pool on big-VRAM cards). Since then a request that stops moving ends with an error instead of hanging (after 2 minutes; 1 minute from 0.1.13): the log says `no progress for ... s ... (issue #29)` with where it stopped, and the next request starts the engine again. If you see that line, please open an issue with it. Engine 0.1.13 adds a stall report under it (what every expert-pool thread and the GPU handshake were doing, memory and page faults) and, on Windows, a `strata-stall-<pid>.dmp` file with every thread's stack: attach both. (`STRATA_WATCHDOG_S` sets the time in seconds; 0 turns it off.) Engine 0.1.14 fixes the stall those reports found (issue #31: with the IQ packs the host could wait forever inside the NVIDIA driver while copying experts in a verify window; the experts are now copied by a GPU kernel, `--pcie-mode dma` restores the old way). |
+| `out of memory: cudaFuncSetAttribute` in the log (IQ3_XXS, long prompt) | Fixed in engine 0.1.15: CUDA loaded a kernel's code when it was first needed, and mid-prompt there was no VRAM left for it. Run `START-HERE.bat` (Windows) or `./setup.sh` (Linux) once to update. |
 | Anything else | The engine log is `strata-<model>.log` in this folder. |
 
 ---
@@ -305,13 +520,15 @@ test images.
 <p align="center"><img src="paper/tiers.svg" width="760" alt="memory tiers"></p>
 
 - **GPU (VRAM):** attention and DeltaNet mixers, the gated-residual weights, routers, shared experts, output head, the MTP
-  draft layer, the KV cache, and an **expert cache** that fills the rest of VRAM with the most-used experts (it adapts to
+  draft layer, the KV cache (from 64K: only its most-read part, the rest streams from RAM), and an **expert cache** that fills the rest of VRAM with the most-used experts (it adapts to
   the conversation while you chat).
 - **RAM:** all 24,576 experts, pinned. The CPU computes the experts that are not on the GPU **in place**, at the same time
   as the GPU works on the cached ones (AVX-512 / AVX2 kernels, ggml's for the i-quants).
 - **SSD:** the 28.8 GB n-gram table, read a few rows per token through the OS cache.
 - **Speculation:** the model's own MTP layer drafts up to 3 tokens; one pass over all 48 layers checks them. 2.4-3.2
-  tokens per pass on average.
+  tokens per pass on average. When the reply repeats the context (code edits, quoted text), **prompt lookup** (engine
+  0.1.7) drafts up to 5 tokens from the earlier copy, but only where its measured acceptance and cost say it pays:
+  code edits 6-11% faster, other text unchanged. The drafts are checked like the MTP's, so the output is the same.
 - **Prompts** are processed in 2,048-token chunks with the experts streamed to the GPU over PCIe.
 
 The full story, with measurements, bottlenecks and what comes next: **[docs/paper/Strata-Paper.pdf](paper/Strata-Paper.pdf)**.
@@ -320,8 +537,12 @@ The full story, with measurements, bottlenecks and what comes next: **[docs/pape
 
 ## Credits and licenses
 
+Strata itself: [MIT](../LICENSE). The model files are not part of it; their licenses apply to them (below).
+
 - Model: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) by the Qwen team; quantizations:
   [ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF).
+  The Coder: [ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF)
+  (Apache-2.0 per its card); its support in Strata came from @pjgmobile's PR #54.
   Swift 1.5: [ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF)
   by UkisAI. Their licenses apply to the weights.
 - [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) (MIT): the i-quant formats, the GPU dot products and
@@ -330,3 +551,7 @@ The full story, with measurements, bottlenecks and what comes next: **[docs/pape
   `third_party/ggml/LICENSE`.
 - Ideas from [Splash](https://github.com/incoai/splash), [ninfer](https://github.com/Neroued/ninfer) and
   [HyperQwen](https://github.com/syv-ai/HyperQwen); references in the paper.
+- The web app's font: [Outfit](https://github.com/Outfitio/Outfit-Fonts) (SIL Open Font License 1.1, see
+  `serve/web/fonts/OFL.txt`). Its Monitor tab started from @code-martin's dashboard idea (PR #22).
+- The experimental speed projection's vector (`data/experimental-speed-projection/`): Qwen Community License 1.0,
+  made from the model's activations (see its README).

docs/MULTI_GPU.md

@@ -0,0 +1,85 @@
+# Strata on two or three GPUs (layer split)
+
+One model can run across several NVIDIA cards in one PC. The layers are split into contiguous ranges, one per GPU:
+the first card runs layers 0 to K-1, the next card runs K onward, and so on; the last card also runs the output head
+and the draft (MTP) layer. Each card keeps an expert cache for **its own layers only**, so two cards hold about twice
+the experts one card holds - for the Coder model on a 16 GB + 24 GB pair, nearly all of them, which is where the
+speed comes from (decode then barely touches the CPU pool).
+
+This is pipeline (layer) parallelism, not tensor parallelism: a token crosses from one card to the next once per
+verify window (a few hundred KB through pinned RAM), not twice per layer. No NVLink or peer-to-peer access is
+needed; cards on x4 or x1 slots work, and the PCIe share of each card is probed on its own link.
+
+## Using it
+
+At setup time (Windows: `START-HERE.bat`, Linux: `./setup.sh`):
+
+```
+--gpus 0,2                 the cards, as nvidia-smi numbers them; the first is the main one
+--layer-split auto         (default) or the first layer of each later card, e.g. 18 or 16,32
+```
+
+or in an existing config (`strata-*.json`), then restart:
+
+```json
+"gpu": [0, 2],
+"layer_split": "auto"
+```
+
+The engine flags behind it: `--layer-split K1[,K2..]|auto` and `--split-device D1[,D2..]` (the later stages'
+devices; default the next visible ones). `--layer-split K --split-device 0` runs both stages on one card sharing
+everything - the bit-exact check of the hand-off, not a speed mode.
+
+**auto** tries every placement (all of them for two or three cards; proportional to the free VRAM beyond that) and
+keeps the one whose caches would hold the most of the expert profile, hottest pairs weighted most; ties go to the
+placement that leaves the fullest card the most room. The startup log prints the choice:
+
+```
+strata generate: layer split auto: K=19 - the caches hold 11767 of 12288 profiled pairs (fullest device 100%)
+strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window
+```
+
+## What each card holds
+
+- **every card**: a copy of the dense weights (~3.4 GB for the Coder), its own session state (the KV cache of the full
+  context), its verify window and its prompt-path buffers, and an expert cache for its layers filled from the profile;
+- **the last card**: also the output head and the draft layer (~0.8 GB);
+- **host RAM**: the expert arena once, shared by all cards (the CPU pool computes whatever no card holds).
+
+Prompts are read in chunks that flow through the cards in turn; while a later card reads chunk c, the first card
+already reads chunk c+1. Conversation checkpoints save and restore every card's state; the adaptive expert swaps copy
+into the card that owns the layer.
+
+## Limits (for now)
+
+- Not with images (`--vision`), control vectors / the speed projection, KV streaming (`--kv-resident`), the older
+  helper-GPU experiment (`--expert-cache-remote`, docs/SECOND_GPU.md) or `--mmap-experts`.
+- The prompt path has its own buffers on every card (1.5 GB each at the default 2048-token chunk; `--prefill 1024`
+  halves that) instead of borrowing cache slots as one card does.
+- Mid-prompt checkpoints (`--prompt-cache-every`) are off; the ones at turn boundaries - what a chat reuses - stay.
+- On Windows only 8 GiB of the expert arena is pinned (more, mapped into two GPU contexts, leaves WDDM refusing
+  allocations); the PCIe share covers those layers.
+- Every card needs compute capability 8.0 (RTX 30 or newer). A Turing card (RTX 20, sm_75) builds only with the
+  experimental `-DSTRATA_EXPERIMENTAL_SM75=ON`.
+
+## Measured
+
+The Coder on an RTX 5080 + RTX 3090 (Ryzen 9 9950X3D), 32K context; details in
+`bench/results/2026-09-29-layer-split/`:
+
+| | Prompt 16K / 28K tok/s | Decode story / code tok/s |
+|---|---|---|
+| 5080 alone | 1,726-2,017 / 1,970 | 83-87 / 88-105 |
+| 5080 + 3090, best split (K=26) | 2,039 / 2,357 | 84 / 110 |
+| 5080 + 3090, auto (K=22) | 2,037 / 2,073 | 80 / 109 |
+
+- **Prompts gain the most** (+18-20%): each card reads its own layers of the chunk while the other reads the next.
+- **Decode is on par with the faster card alone**, and ahead on code. Once both caches hold nearly every routed
+  expert, the per-layer GPU time decides.
+- **Correctness:** one GPU is byte-identical to 0.1.20, and the hand-off itself is bit-exact.
+
+**Which cards and in what order:**
+- Put the fastest card first; auto gives it as many layers as its cache allows.
+- Leave out a much slower card when two already hold the model. An RTX 2080 Ti as a third card made the 5080 +
+  3090 pair slower (68 / 90 tok/s decode): every extra card costs its own round per window.
+- More cards pay off when the model's routed experts do not fit the faster ones.

docs/ORCA.md

@@ -0,0 +1,95 @@
+# OrcaRouter IQ3_XXS compatibility
+
+Validated on Linux with an RTX 5090 (32 GB), Ryzen 9 9950X3D and 128 GB RAM on 2026-09-27.
+
+The target is `orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF`, **IQ3_XXS** (two shards, 85.20 GB total).
+It has Strata's Qwen4Exp geometry, but ordinary quantization also compresses the hyperconnection,
+small attention and PLE projections which Strata's kernels read as BF16.
+
+`tools/iq_pack.py --compat-bf16` expands only those projections, rounding to BF16 using nearest-even.
+This adds rounding relative to the GGUF's dequantized weights; it does not recover the original BF16
+checkpoint or its quality. Expert weights, native attention weights, embeddings and the 28.8 GB PLE
+lookup table remain unchanged. The latter stays on disk. The pack records its conversions in
+`compat-bf16.json`.
+
+The native loader retains this pack's converted PLE key. Its existing native Q2_0 key path remains
+available to the original model. Hugging Face snapshot symlinks are supported; all split shards must
+finish downloading before packing. Pass the snapshot filename, not its hash-named blob target.
+
+## Preparation
+
+Build Strata normally using the pinned llama.cpp dependency. Python needs numpy, regex and gguf-py
+from that dependency (`STRATA_GGUF_PY` can point to its `gguf-py` directory).
+
+```sh
+.venv/bin/python tools/iq_pack.py \
+  --gguf /path/to/Qwen3.8-Flash-Next-Uncensored-IQ3_XXS-00001-of-00002.gguf \
+  --out packs/orca-iq3_xxs --compat-bf16
+```
+
+Use this model's own tokenizer exported into the pack. Do not share another model's `dense.bin`,
+and do not rename the Orca files to impersonate one of setup's GSQ-RCO models.
+
+Strata's persistent server also requires the MTP runtime, prepared by the existing tools:
+
+```sh
+.venv/bin/python tools/mtp_fetch.py fetch --out mtp
+.venv/bin/python tools/mtp_pack.py --src mtp --experts q2_0 --out mtp/mtp-q2_0.gguf
+.venv/bin/python tools/mtp_rt.py --gguf mtp/mtp-q2_0.gguf --out mtp/rt
+cp data/draft_vocab.bin mtp/rt/draft_vocab.bin
+```
+
+This uses the original model's draft head, as setup does for Swift. The target model verifies its
+proposals; draft acceptance and performance must be measured for this fine-tune.
+
+## Local server
+
+Save the following as `strata-orca-iq3_xxs.json` at the repository root, replacing both `/path/to/`
+entries with the **same first shard**. Orca's PLE table is in shard 1. Start with a 32K context and
+512-token prefill chunks; the expert cache sizes itself to available VRAM. The expert arena alone
+needs about 49.8 GiB of available system RAM, plus draft/runtime buffers and other applications.
+
+```json
+{
+  "exe": "build/strata",
+  "args": [
+    "--pack", "packs/orca-iq3_xxs",
+    "--native", "/path/to/Qwen3.8-Flash-Next-Uncensored-IQ3_XXS-00001-of-00002.gguf",
+    "--ple-gguf", "/path/to/Qwen3.8-Flash-Next-Uncensored-IQ3_XXS-00001-of-00002.gguf",
+    "--expert-profile", "data/expert-profile.bin", "--expert-cache", "auto",
+    "--prefill", "512", "--spec", "4", "--spec-min-p", "0.5",
+    "--mtp", "mtp/rt", "--max-context", "32768", "--kv", "int8"
+  ],
+  "cwd": ".",
+  "tokenizer": "packs/orca-iq3_xxs/tokenizer",
+  "model_name": "orcarouter-qwen3.8-flash-next-uncensored-iq3_xxs",
+  "log": "strata-orca-iq3_xxs.log",
+  "host": "127.0.0.1",
+  "port": 8080
+}
+```
+
+Run from the repository root:
+
+```sh
+.venv/bin/python -m serve.server --engine strata --config strata-orca-iq3_xxs.json --port 8080
+```
+
+The web interface is at `http://127.0.0.1:8080`; API clients use `http://127.0.0.1:8080/v1`.
+Vision is not configured in this text-only example. Performance of the original GSQ-RCO model
+does not establish this fine-tune's speed or accuracy.
+
+## Scope and validation
+
+- IQ3_XXS is the target of this change; other Orca quantizations are not validated.
+- IQ3_M additionally uses Q5_0 expert down matrices, which the native GPU expert path does not support.
+- The installer model menu is unchanged. This is an explicit local packing workflow.
+- Focused conversion and split-file tests: `.venv/bin/python -m unittest discover -s tools -p test_iq_pack.py`.
+- All eight packing tests, 17 server tests and the GPU gated-residual parity check passed.
+- The complete model packed successfully: 460 tensors converted, 1.39 GiB of converted BF16 weights;
+  1.43 GiB total packed dense weights. The 48 expert layers remain IQ3_XXS gate/up and IQ4_NL down.
+- Real server checks passed: arithmetic, Python code, translation, a longer explanation, multi-turn context
+  reuse (25 prompt tokens reused), streaming cancellation and a subsequent correct response. The engine
+  PID stayed unchanged. These are smoke tests, not a quality benchmark or a full-context stress test.
+- With the configuration above, a 111-token explanation generated at 77.7 tokens/s (engine decode timing);
+  total request time was 1.77 seconds. This single short measurement is not a general throughput claim.

docs/SECOND_GPU.md

@@ -0,0 +1,84 @@
+# CUDA1–3 expert experiment
+
+CUDA0 keeps the dense layers, KV/state, MTP and its existing expert cache. Up to
+three independent expert caches fill CUDA1, CUDA2 and CUDA3 in that order. Each
+expert has one owner: CUDA0, one secondary GPU, or the CPU pool. Each secondary
+GPU computes its rows while the CPU pool works; the results return through
+private pinned host buffers before CUDA0 continues the layer. The GPU launches
+can overlap, but each layer waits for all three.
+
+The shipped profile ranks 8,000 `(layer, expert)` pairs. CUDA0 takes its usual
+`--expert-cache auto` allocation. With only CUDA1, it takes the first remaining
+pairs. When CUDA2 or CUDA3 is enabled, the remaining profiled pairs are assigned
+round robin to the secondary GPUs. The ranked list is then extended with all
+other pairs in expert-then-layer order and distributed in the same way. This
+fills the additional VRAM and lets all secondary GPUs see frequent experts, but
+unranked experts may hardly ever be routed. The startup
+log reports slots and GiB per card; each request reports the number of expert
+entries actually computed on each secondary GPU. Remote outputs are packed
+before returning over PCIe/USB4: the request log compares MiB transferred
+with the previous full-row transfer. The input and routing metadata use pinned
+host staging. This reduces transfer traffic and launch overhead, but does not
+release the expert arena in system RAM or significantly change VRAM use.
+
+The optional `--gpu-placement layer` gives each layer to exactly one secondary
+GPU (layer number modulo the number of enabled cards). It fills that card with
+the most frequently routed experts of its layers first, then the unranked ones.
+This can reduce USB4 device switches and synchronization when several secondary
+cards serve the same layer in the default `stripe` mode. All three cards still
+keep their requested expert slots; the primary CUDA0 cache and CPU fallback
+continue to work. The log counts active layer launches on each card, so the
+tradeoff can be measured alongside tokens per second. This placement is
+experimental and may be slower if a given layer needs more than one card's
+compute capacity.
+
+The experiment requires the corresponding visible CUDA devices and a CUDA
+enabled build. Each card must keep at least 512 MiB free. Tiers must be enabled
+in order. On Windows, CUDA registration of the host expert arena is capped at
+8 GiB to leave room for the contexts and MTP on CUDA0. The rest remains
+available to the CPU pool; the PCIe expert path is available only for the
+registered layers. Without secondary GPUs, the original uncapped registration
+behavior applies.
+
+On an existing Linux installation, apply the patch to its source tree, then run:
+
+```sh
+./setup.sh --setup --gpu1-experts 5000 --gpu2-experts 5000 --gpu3-experts 5000
+```
+
+On Windows, from PowerShell in the installation folder:
+
+```powershell
+.\START-HERE.bat --setup --gpu1-experts 5000 --gpu2-experts 5000 --gpu3-experts 5000
+```
+
+To test the layer placement instead, add `--gpu-placement layer` to the setup
+command. To go back, run the same command with `--gpu-placement stripe`. After
+the new engine is built, the mode can also be changed without a rebuild: edit
+`--expert-cache-remote-placement` in `strata-iq3_xxs.json` from `layer` to
+`stripe` or vice versa and restart the server. If the argument is absent, the
+default is `stripe`.
+
+Choose the same model, size, context and image settings as your current
+installation when prompted. Setup recompiles the engine, updates its server
+configuration, and starts it. Later starts keep those settings. To tune the
+slots without recompiling, edit the values after `--expert-cache-device1`,
+`--expert-cache-device2` and `--expert-cache-device3` in the stored
+`strata-iq3_xxs.json`, then restart with `.\run-iq3_xxs.bat`.
+
+Compare identical requests at one, two and four GPUs, preferably with several
+repeats. VRAM use alone does not show useful offload: compare the per-request
+CUDA1–3 counts and tokens per second. More GPU contexts and synchronization
+may lower the speed. Prompt prefill still uses CUDA0; secondary GPUs serve
+decode, including MTP verification. No peer-to-peer access is required.
+
+Longer contexts reserve more KV/state memory on CUDA0, which reduces its
+automatic expert cache. On a 64 GB PC `setup.py` limits IQ3_XXS to 128K even
+if 262K was selected; the secondary caches still require the host expert arena.
+To try 262K on 64 GB, choose Q2_0 or IQ2_XS instead and check that CUDA0 still
+has enough free VRAM after loading. Benchmark long prompts separately from
+short decode requests.
+
+The existing warning about the CUDA0 expert-cache GPU hit path still applies:
+its outputs diverge from cache-off runs. Treat performance as experimental
+until the generated tokens have been validated.

docs/media/guess-and-check.svg

@@ -0,0 +1,23 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 960 330" width="960" height="330" role="img" aria-label="Speculative decoding: a small helper guesses words, the big model checks them at once"><title>Speculative decoding: a small helper guesses words, the big model checks them at once</title><style>
+@font-face{font-family:"Outfit";src:url(data:font/woff2;base64,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 truncated
+text{font-family:"Outfit",system-ui,-apple-system,"Segoe UI",Roboto,sans-serif;fill:#221e1f}
+.bg{fill:#f9fafb}.card{fill:#ffffff;stroke:#dee5ee;stroke-width:1}.soft{fill:#f7f7f7}
+.muted{fill:#6f6f6f}.ink{fill:#221e1f}.icon{color:#2F384C}
+.acc{fill:#10b981}.acc-t{fill:#0c8a60}.acc-tint{fill:#ecfdf5}.acc-line{stroke:#10b981}
+.info{fill:#0f97ff}.info-t{fill:#0668c2}.info-tint{fill:#e8f4ff}
+.warn{fill:#e39b0b}.warn-t{fill:#8a5a00}.warn-tint{fill:#fff6e0}
+.dan{fill:#e0283f}.dan-tint{fill:#fdecee}.dan-t{fill:#b01a2e}
+.dot{fill:#dee5ee}.line{stroke:#dee5ee}.arrow{stroke:#6f6f6f;fill:none}.arrowhead{fill:#6f6f6f}
+.dash{stroke:#6f6f6f;fill:none;stroke-dasharray:4 4}
+.on-acc{fill:#ffffff}.icon-acc{color:#0c8a60}.icon-dan{color:#b01a2e}
+@media (prefers-color-scheme:dark){
+text,.ink{fill:#eef1f3}.bg{fill:#0e1113}.card{fill:#161a1d;stroke:#2a3036}.soft{fill:#1d2226}
+.muted{fill:#8e979f}.icon{color:#c7cdd3}
+.acc{fill:#34d399}.acc-t{fill:#4ade80}.acc-tint{fill:rgba(52,211,153,.12)}.acc-line{stroke:#34d399}
+.info{fill:#3aa9ff}.info-t{fill:#7cc4ff}.info-tint{fill:rgba(58,169,255,.12)}
+.warn{fill:#f5b53d}.warn-t{fill:#f8c96a}.warn-tint{fill:rgba(245,181,61,.12)}
+.dan{fill:#ff5a6e}.dan-tint{fill:rgba(255,90,110,.12)}.dan-t{fill:#ff8a98}
+.dot{fill:#2a3036}.line{stroke:#2a3036}.arrow{stroke:#8e979f}.arrowhead{fill:#8e979f}.dash{stroke:#8e979f}
+.on-acc{fill:#052e20}.icon-acc{color:#4ade80}.icon-dan{color:#ff8a98}
+}
+</style><rect x="0" y="0" width="960" height="330" rx="20" class="bg"/><rect x="24" y="24" width="912" height="282" rx="14" class="card"/><g class="icon" transform="translate(48,48) scale(1.2500)"><path opacity="0.12" d="M12 2.5C8.41 2.5 5.5 5.41 5.5 9C5.5 11.3 6.7 13.3 8.5 14.5V16.5C8.5 17.05 8.95 17.5 9.5 17.5H14.5C15.05 17.5 15.5 17.05 15.5 16.5V14.5C17.3 13.3 18.5 11.3 18.5 9C18.5 5.41 15.59 2.5 12 2.5Z" fill="currentColor"></path><path d="M12 2.5C8.41 2.5 5.5 5.41 5.5 9C5.5 11.3 6.7 13.3 8.5 14.5V16.5C8.5 17.05 8.95 17.5 9.5 17.5H14.5C15.05 17.5 15.5 17.05 15.5 16.5V14.5C17.3 13.3 18.5 11.3 18.5 9C18.5 5.41 15.59 2.5 12 2.5Z" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path><path d="M9.5 20.5H14.5" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><text x="92" y="70" font-size="19" font-weight="700" class="ink" text-anchor="start">Guess, then check - many words per step</text><text x="92" y="94" font-size="15" font-weight="400" class="muted" text-anchor="start">A small, fast helper inside the model guesses the next words. The big model checks all the guesses in one go</text><text x="92" y="113" font-size="15" font-weight="400" class="muted" text-anchor="start">and keeps the ones it agrees with. The big model decides every word - it just arrives 1.6-1.8x sooner.</text><text x="48" y="180" font-size="14" font-weight="600" class="muted" text-anchor="start">Helper guesses</text><text x="48" y="262" font-size="14" font-weight="600" class="muted" text-anchor="start">Big model checks</text><rect x="176" y="156" width="61" height="38" rx="10" class="dash" stroke-width="1.4"/><text x="206.5" y="181" font-size="16" font-weight="500" class="ink" text-anchor="middle">def</text><rect x="251" y="156" width="97" height="38" rx="10" class="dash" stroke-width="1.4"/><text x="299.5" y="181" font-size="16" font-weight="500" class="ink" text-anchor="middle">reverse</text><rect x="362" y="156" width="88" height="38" rx="10" class="dash" stroke-width="1.4"/><text x="406.0" y="181" font-size="16" font-weight="500" class="ink" text-anchor="middle">(items</text><rect x="464" y="156" width="52" height="38" rx="10" class="dash" stroke-width="1.4"/><text x="490.0" y="181" font-size="16" font-weight="500" class="ink" text-anchor="middle">):</text><rect x="176" y="238" width="61" height="38" rx="10" class="acc-tint"/><text x="198.5" y="263" font-size="16" font-weight="600" class="acc-t" text-anchor="middle">def</text><g class="icon-acc" transform="translate(215,249) scale(0.6667)"><path d="M4.5 12.5L9.5 17.5L19.5 6.5" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><line x1="206.5" y1="196" x2="206.5" y2="234" class="line" stroke-width="1.4"/><rect x="251" y="238" width="97" height="38" rx="10" class="acc-tint"/><text x="291.5" y="263" font-size="16" font-weight="600" class="acc-t" text-anchor="middle">reverse</text><g class="icon-acc" transform="translate(326,249) scale(0.6667)"><path d="M4.5 12.5L9.5 17.5L19.5 6.5" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><line x1="299.5" y1="196" x2="299.5" y2="234" class="line" stroke-width="1.4"/><rect x="362" y="238" width="88" height="38" rx="10" class="acc-tint"/><text x="398.0" y="263" font-size="16" font-weight="600" class="acc-t" text-anchor="middle">(items</text><g class="icon-acc" transform="translate(428,249) scale(0.6667)"><path d="M4.5 12.5L9.5 17.5L19.5 6.5" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><line x1="406.0" y1="196" x2="406.0" y2="234" class="line" stroke-width="1.4"/><rect x="464" y="238" width="52" height="38" rx="10" class="dan-tint"/><text x="482.0" y="263" font-size="16" font-weight="600" class="dan-t" text-anchor="middle">):</text><g class="icon-dan" transform="translate(494,249) scale(0.6667)"><path opacity="0.12" d="M22 12C22 17.5228 17.5228 22 12 22C6.47715 22 2 17.5228 2 12C2 6.47715 6.47715 2 12 2C17.5228 2 22 6.47715 22 12Z" fill="currentColor"></path><path d="M22 12C22 17.5228 17.5228 22 12 22C6.47715 22 2 17.5228 2 12C2 6.47715 6.47715 2 12 2C17.5228 2 22 6.47715 22 12Z" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path><path d="M15 9L9 15M9 9L15 15" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><rect x="526" y="238" width="89" height="38" rx="10" class="acc-tint"/><text x="570.5" y="263" font-size="16" font-weight="600" class="acc-t" text-anchor="middle">, key):</text><line x1="490.0" y1="196" x2="490.0" y2="234" class="line" stroke-width="1.4"/><line x1="655" y1="257" x2="678" y2="257" class="arrow" stroke-width="1.6" stroke-linecap="round"/><path d="M685 257 l-8 -5 v10 z" class="arrowhead"/><rect x="697" y="232" width="215" height="50" rx="12" class="acc"/><g class="icon" style="color:#ffffff" transformLine truncated

docs/media/how-it-works.svg

@@ -0,0 +1,23 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 960 470" width="960" height="470" role="img" aria-label="How Strata splits the model between your graphics card, RAM and SSD"><title>How Strata splits the model between your graphics card, RAM and SSD</title><style>
+@font-face{font-family:"Outfit";src:url(data:font/woff2;base64,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 truncated
+text{font-family:"Outfit",system-ui,-apple-system,"Segoe UI",Roboto,sans-serif;fill:#221e1f}
+.bg{fill:#f9fafb}.card{fill:#ffffff;stroke:#dee5ee;stroke-width:1}.soft{fill:#f7f7f7}
+.muted{fill:#6f6f6f}.ink{fill:#221e1f}.icon{color:#2F384C}
+.acc{fill:#10b981}.acc-t{fill:#0c8a60}.acc-tint{fill:#ecfdf5}.acc-line{stroke:#10b981}
+.info{fill:#0f97ff}.info-t{fill:#0668c2}.info-tint{fill:#e8f4ff}
+.warn{fill:#e39b0b}.warn-t{fill:#8a5a00}.warn-tint{fill:#fff6e0}
+.dan{fill:#e0283f}.dan-tint{fill:#fdecee}.dan-t{fill:#b01a2e}
+.dot{fill:#dee5ee}.line{stroke:#dee5ee}.arrow{stroke:#6f6f6f;fill:none}.arrowhead{fill:#6f6f6f}
+.dash{stroke:#6f6f6f;fill:none;stroke-dasharray:4 4}
+.on-acc{fill:#ffffff}.icon-acc{color:#0c8a60}.icon-dan{color:#b01a2e}
+@media (prefers-color-scheme:dark){
+text,.ink{fill:#eef1f3}.bg{fill:#0e1113}.card{fill:#161a1d;stroke:#2a3036}.soft{fill:#1d2226}
+.muted{fill:#8e979f}.icon{color:#c7cdd3}
+.acc{fill:#34d399}.acc-t{fill:#4ade80}.acc-tint{fill:rgba(52,211,153,.12)}.acc-line{stroke:#34d399}
+.info{fill:#3aa9ff}.info-t{fill:#7cc4ff}.info-tint{fill:rgba(58,169,255,.12)}
+.warn{fill:#f5b53d}.warn-t{fill:#f8c96a}.warn-tint{fill:rgba(245,181,61,.12)}
+.dan{fill:#ff5a6e}.dan-tint{fill:rgba(255,90,110,.12)}.dan-t{fill:#ff8a98}
+.dot{fill:#2a3036}.line{stroke:#2a3036}.arrow{stroke:#8e979f}.arrowhead{fill:#8e979f}.dash{stroke:#8e979f}
+.on-acc{fill:#052e20}.icon-acc{color:#4ade80}.icon-dan{color:#ff8a98}
+}
+</style><rect x="0" y="0" width="960" height="470" rx="20" class="bg"/><rect x="24" y="24" width="912" height="92" rx="14" class="card"/><g class="icon" transform="translate(48,50) scale(1.5000)"><path opacity="0.12" d="M21 11.5C21 16.194 16.971 20 12 20C10.8 20 9.65 19.78 8.6 19.37L4 20.5L5.2 16.8C3.83 15.34 3 13.5 3 11.5C3 6.806 7.029 3 12 3C16.971 3 21 6.806 21 11.5Z" fill="currentColor"></path><path d="M21 11.5C21 16.194 16.971 20 12 20C10.8 20 9.65 19.78 8.6 19.37L4 20.5L5.2 16.8C3.83 15.34 3 13.5 3 11.5C3 6.806 7.029 3 12 3C16.971 3 21 6.806 21 11.5Z" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path><path d="M8 11.5H8.01M12 11.5H12.01M16 11.5H16.01" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round"></path></g><text x="100" y="62" font-size="19" font-weight="600" class="ink" text-anchor="start">The model is made of 24,576 small specialists - “experts”.</text><text x="100" y="88" font-size="15" font-weight="400" class="muted" text-anchor="start">For each word it writes, it asks only 10 of them. Strata keeps the busy ones</text><text x="100" y="106" font-size="15" font-weight="400" class="muted" text-anchor="start">on your graphics card and the rest in RAM - so it runs on a normal PC.</text><circle cx="698.0" cy="46.0" r="2.6" class="dot"/><circle cx="705.6" cy="46.0" r="2.6" class="dot"/><circle cx="713.2" cy="46.0" r="2.6" class="dot"/><circle cx="720.8" cy="46.0" r="2.6" class="dot"/><circle cx="728.4" cy="46.0" r="2.6" class="dot"/><circle cx="736.0" cy="46.0" r="2.6" class="dot"/><circle cx="743.6" cy="46.0" r="2.6" class="dot"/><circle cx="751.2" cy="46.0" r="2.6" class="dot"/><circle cx="758.8" cy="46.0" r="2.6" class="acc"/><circle cx="766.4" cy="46.0" r="2.6" class="dot"/><circle cx="774.0" cy="46.0" r="2.6" class="dot"/><circle cx="781.6" cy="46.0" r="2.6" class="dot"/><circle cx="789.2" cy="46.0" r="2.6" class="dot"/><circle cx="796.8" cy="46.0" r="2.6" class="dot"/><circle cx="804.4" cy="46.0" r="2.6" class="dot"/><circle cx="812.0" cy="46.0" r="2.6" class="dot"/><circle cx="819.6" cy="46.0" r="2.6" class="acc"/><circle cx="827.2" cy="46.0" r="2.6" class="dot"/><circle cx="834.8" cy="46.0" r="2.6" class="dot"/><circle cx="842.4" cy="46.0" r="2.6" class="dot"/><circle cx="850.0" cy="46.0" r="2.6" class="dot"/><circle cx="857.6" cy="46.0" r="2.6" class="dot"/><circle cx="865.2" cy="46.0" r="2.6" class="dot"/><circle cx="872.8" cy="46.0" r="2.6" class="dot"/><circle cx="880.4" cy="46.0" r="2.6" class="dot"/><circle cx="888.0" cy="46.0" r="2.6" class="dot"/><circle cx="895.6" cy="46.0" r="2.6" class="dot"/><circle cx="903.2" cy="46.0" r="2.6" class="dot"/><circle cx="698.0" cy="53.6" r="2.6" class="dot"/><circle cx="705.6" cy="53.6" r="2.6" class="dot"/><circle cx="713.2" cy="53.6" r="2.6" class="acc"/><circle cx="720.8" cy="53.6" r="2.6" class="dot"/><circle cx="728.4" cy="53.6" r="2.6" class="dot"/><circle cx="736.0" cy="53.6" r="2.6" class="dot"/><circle cx="743.6" cy="53.6" r="2.6" class="dot"/><circle cx="751.2" cy="53.6" r="2.6" class="dot"/><circle cx="758.8" cy="53.6" r="2.6" class="dot"/><circle cx="766.4" cy="53.6" r="2.6" class="dot"/><circle cx="774.0" cy="53.6" r="2.6" class="dot"/><circle cx="781.6" cy="53.6" r="2.6" class="dot"/><circle cx="789.2" cy="53.6" r="2.6" class="dot"/><circle cx="796.8" cy="53.6" r="2.6" class="dot"/><circle cx="804.4" cy="53.6" r="2.6" class="dot"/><circle cx="812.0" cy="53.6" r="2.6" class="dot"/><circle cx="819.6" cy="53.6" r="2.6" class="dot"/><circle cx="827.2" cy="53.6" r="2.6" class="dot"/><circle cx="834.8" cy="53.6" r="2.6" class="dot"/><circle cx="842.4" cy="53.6" r="2.6" class="acc"/><circle cx="850.0" cy="53.6" r="2.6" class="dot"/><circle cx="857.6" cy="53.6" r="2.6" class="dot"/><circle cx="865.2" cy="53.6" r="2.6" class="dot"/><circle cx="872.8" cy="53.6" r="2.6" class="dot"/><circle cx="880.4" cy="53.6" r="2.6" class="dot"/><circle cx="888.0" cy="53.6" r="2.6" class="dot"/><circle cx="895.6" cy="53.6" r="2.6" class="dot"/><circle cx="903.2" cy="53.6" r="2.6" class="dot"/><circle cx="698.0" cy="61.2" r="2.6" class="dot"/><circle cx="705.6" cy="61.2" r="2.6" class="dot"/><circle cx="713.2" cy="61.2" r="2.6" class="dot"/><circle cx="720.8" cy="61.2" r="2.6" class="dot"/><circle cx="728.4" cy="61.2" r="2.6" class="acc"/><circle cx="736.0" cy="61.2" r="2.6" class="dot"/><circle cx="743.6" cy="61.2" r="2.6" class="dot"/><circle cx="751.2" cy="61.2" r="2.6" class="dot"/><circle cx="758.8" cy="61.2" r="2.6" class="dot"/><circle cx="766.4" cy="61.2" r="2.6" class="dot"/><circle cx="774.0" cy="61.2" r="2.6" class="dot"/><circle cx="781.6" cy="61.2" r="2.6" class="dot"/><circle cx="789.2" cy="61.2" r="2.6" class="dot"/><circle cx="796.8" cy="61.2" r="2.6" class="acc"/><circle cx="804.4" cy="61.2" r="2.6" class="dot"/><circle cx="812.0" cy="61.2" r="2.6" class="dot"/><circle cx="819.6" cy="61.2" r="2.6" class="dot"/><circle cx="827.2" cy="61.2" r="2.6" class="dot"Line truncated

docs/media/make_figures.py

@@ -0,0 +1,178 @@
+"""docs/media/make_figures.py - the README's "How does it work?" illustrations, in the Strata app's style: its colors
+(serve/web/tokens.css), its icons (serve/web/sprite.svg) and its font (Outfit, embedded, SIL OFL 1.1).  Light and dark
+follow the reader's system setting.
+
+    python docs/media/make_figures.py        # writes docs/media/how-it-works.svg and docs/media/guess-and-check.svg
+"""
+import base64
+import random
+import re
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+OUT = ROOT / "docs" / "media"
+SPRITE = (ROOT / "serve" / "web" / "sprite.svg").read_text(encoding="utf-8")
+FONT = base64.b64encode((ROOT / "serve" / "web" / "fonts" / "outfit-latin-wght.woff2").read_bytes()).decode()
+
+# the app's tokens (light, then dark)
+CSS = """
+@font-face{font-family:"Outfit";src:url(data:font/woff2;base64,%s) format("woff2");font-weight:100 900}
+text{font-family:"Outfit",system-ui,-apple-system,"Segoe UI",Roboto,sans-serif;fill:#221e1f}
+.bg{fill:#f9fafb}.card{fill:#ffffff;stroke:#dee5ee;stroke-width:1}.soft{fill:#f7f7f7}
+.muted{fill:#6f6f6f}.ink{fill:#221e1f}.icon{color:#2F384C}
+.acc{fill:#10b981}.acc-t{fill:#0c8a60}.acc-tint{fill:#ecfdf5}.acc-line{stroke:#10b981}
+.info{fill:#0f97ff}.info-t{fill:#0668c2}.info-tint{fill:#e8f4ff}
+.warn{fill:#e39b0b}.warn-t{fill:#8a5a00}.warn-tint{fill:#fff6e0}
+.dan{fill:#e0283f}.dan-tint{fill:#fdecee}.dan-t{fill:#b01a2e}
+.dot{fill:#dee5ee}.line{stroke:#dee5ee}.arrow{stroke:#6f6f6f;fill:none}.arrowhead{fill:#6f6f6f}
+.dash{stroke:#6f6f6f;fill:none;stroke-dasharray:4 4}
+.on-acc{fill:#ffffff}.icon-acc{color:#0c8a60}.icon-dan{color:#b01a2e}
+@media (prefers-color-scheme:dark){
+text,.ink{fill:#eef1f3}.bg{fill:#0e1113}.card{fill:#161a1d;stroke:#2a3036}.soft{fill:#1d2226}
+.muted{fill:#8e979f}.icon{color:#c7cdd3}
+.acc{fill:#34d399}.acc-t{fill:#4ade80}.acc-tint{fill:rgba(52,211,153,.12)}.acc-line{stroke:#34d399}
+.info{fill:#3aa9ff}.info-t{fill:#7cc4ff}.info-tint{fill:rgba(58,169,255,.12)}
+.warn{fill:#f5b53d}.warn-t{fill:#f8c96a}.warn-tint{fill:rgba(245,181,61,.12)}
+.dan{fill:#ff5a6e}.dan-tint{fill:rgba(255,90,110,.12)}.dan-t{fill:#ff8a98}
+.dot{fill:#2a3036}.line{stroke:#2a3036}.arrow{stroke:#8e979f}.arrowhead{fill:#8e979f}.dash{stroke:#8e979f}
+.on-acc{fill:#052e20}.icon-acc{color:#4ade80}.icon-dan{color:#ff8a98}
+}
+""" % FONT
+
+
+def icon(name, x, y, size=28, cls="icon"):
+    """A sprite icon at (x, y), `size` px, drawn in the class's color (the icons use currentColor)."""
+    m = re.search(r'<symbol id="%s"[^>]*>(.*?)</symbol>' % name, SPRITE, re.S)
+    return (f'<g class="{cls}" transform="translate({x},{y}) scale({size / 24:.4f})">{m.group(1)}</g>')
+
+
+def text(x, y, s, size=14, weight=400, cls="ink", anchor="start"):
+    return (f'<text x="{x}" y="{y}" font-size="{size}" font-weight="{weight}" class="{cls}" '
+            f'text-anchor="{anchor}">{s}</text>')
+
+
+def rect(x, y, w, h, r=14, cls="card"):
+    return f'<rect x="{x}" y="{y}" width="{w}" height="{h}" rx="{r}" class="{cls}"/>'
+
+
+def svg(w, h, body, title):
+    return (f'<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 {w} {h}" width="{w}" height="{h}" role="img" '
+            f'aria-label="{title}"><title>{title}</title><style>{CSS}</style>'
+            f'{rect(0, 0, w, h, 20, "bg")}{body}</svg>')
+
+
+def dots(x, y, cols, rows, step, lit, lit_cls="acc", r=2.6, seed=1):
+    """A grid of expert dots; `lit` of them (spread at random) in the accent color."""
+    rnd = random.Random(seed)
+    n = cols * rows
+    on = set(rnd.sample(range(n), min(lit, n)))
+    out = []
+    for i in range(n):
+        cx, cy = x + (i % cols) * step, y + (i // cols) * step
+        out.append(f'<circle cx="{cx:.1f}" cy="{cy:.1f}" r="{r}" class="{lit_cls if i in on else "dot"}"/>')
+    return "".join(out)
+
+
+def arrow(x1, y1, x2, y2, cls="arrow"):
+    return (f'<line x1="{x1}" y1="{y1}" x2="{x2 - 7}" y2="{y2}" class="{cls}" stroke-width="1.6" stroke-linecap="round"/>'
+            f'<path d="M{x2} {y2} l-8 -5 v10 z" class="arrowhead"/>')
+
+
+def how_it_works():
+    W, H = 960, 470
+    b = []
+    # the headline strip: one word, 10 experts
+    b.append(rect(24, 24, W - 48, 92))
+    b.append(icon("i-chat", 48, 50, 36))
+    b.append(text(100, 62, "The model is made of 24,576 small specialists - “experts”.", 19, 600))
+    b.append(text(100, 88, "For each word it writes, it asks only 10 of them. Strata keeps the busy ones",
+                  15, 400, "muted"))
+    b.append(text(100, 106, "on your graphics card and the rest in RAM - so it runs on a normal PC.", 15, 400, "muted"))
+    b.append(dots(W - 262, 46, 28, 4, 7.6, 10, seed=3))
+    b.append(text(W - 262, 94, "10 at work for one word", 12, 500, "acc-t"))
+    # three cards
+    cw, ch, cy = 288, 316, 136
+    xs = [24, 24 + cw + 24, 24 + 2 * (cw + 24)]
+    cards = [
+        ("i-gpu", "Graphics card", "Small but very fast: 12-24 GB", "acc",
+         ["Runs the core of the model", "for every word, and keeps the", "~4,000 experts used most"],
+         "the most-used experts", 5, 26, 26 * 5, "acc"),
+        ("i-memory", "RAM + processor", "Big: 32-64 GB of memory", "info",
+         ["Holds all 24,576 experts. When a", "word needs one the card doesn't", "have, the CPU does it - at once"],
+         "all 24,576 experts", 10, 26, 18, "info"),
+        ("i-disk", "SSD", "Holds a 29 GB lookup table", "warn",
+         ["The model looks up a few", "small rows of it per word -", "no need to keep it in RAM"],
+         "", 0, 0, 0, "warn"),
+    ]
+    for (ic, title, sub, tone, lines, dlabel, rows, cols, lit, lit_cls), x in zip(cards, xs):
+        b.append(rect(x, cy, cw, ch))
+        b.append(rect(x + 20, cy + 20, 48, 48, 12, tone + "-tint"))
+        b.append(icon(ic, x + 30, cy + 30, 28, "icon"))
+        b.append(text(x + 82, cy + 42, title, 19, 700))
+        b.append(text(x + 82, cy + 62, sub, 13, 500, tone + "-t"))
+        for i, ln in enumerate(lines):
+            b.append(text(x + 20, cy + 104 + i * 21, ln, 15, 400, "ink"))
+        if rows:
+            b.append(dots(x + 26, cy + 196, cols, rows, 9.4, lit, lit_cls, seed=len(title)))
+            b.append(text(x + 20, cy + 196 + rows * 9.4 + 14, dlabel, 12, 500, "muted"))
+        else:                                       # the SSD: a table of rows, a few being read
+            for r in range(6):
+                yy = cy + 190 + r * 17
+                b.append(rect(x + 20, yy, cw - 40, 11, 4, "warn" if r in (1, 4) else "soft"))
+            b.append(text(x + 20, cy + 306, "a few rows per word", 12, 500, "muted"))
+    return svg(W, H, "".join(b), "How Strata splits the model between your graphics card, RAM and SSD")
+
+
+def guess_and_check():
+    W, H = 960, 330
+    b = [rect(24, 24, W - 48, H - 48)]
+    b.append(icon("i-thinking", 48, 48, 30))
+    b.append(text(92, 70, "Guess, then check - many words per step", 19, 700))
+    b.append(text(92, 94, "A small, fast helper inside the model guesses the next words. The big model checks all the "
+                          "guesses in one go", 15, 400, "muted"))
+    b.append(text(92, 113, "and keeps the ones it agrees with. The big model decides every word - it just arrives 1.6-1.8x sooner.",
+                  15, 400, "muted"))
+    words = ["def", "reverse", "(items", "):"]
+    fixed = ", key):"                                  # the model's own word where the guess was wrong
+    ok = [True, True, True, False]
+    # row 1: the helper's guesses (dashed chips)
+    y1, y2 = 156, 238
+    b.append(text(48, y1 + 24, "Helper guesses", 14, 600, "muted"))
+    b.append(text(48, y2 + 24, "Big model checks", 14, 600, "muted"))
+    x = 176
+    xs = []
+    for i, w in enumerate(words[:4]):
+        wd = 34 + len(w) * 9
+        xs.append((x, wd))
+        b.append(f'<rect x="{x}" y="{y1}" width="{wd}" height="38" rx="10" class="dash" stroke-width="1.4"/>')
+        b.append(text(x + wd / 2, y1 + 25, w, 16, 500, "ink", "middle"))
+        x += wd + 14
+    # row 2: the verdicts
+    for i, ((x, wd), good) in enumerate(zip(xs, ok)):
+        tone = "acc" if good else "dan"
+        b.append(rect(x, y2, wd, 38, 10, tone + "-tint"))
+        b.append(text(x + wd / 2 - 8, y2 + 25, words[i], 16, 600, tone + "-t", "middle"))
+        b.append(icon("i-check" if good else "i-error", x + wd - 22, y2 + 11, 16, "icon-" + tone))
+        if not good:                                   # the model's own word beside the rejected guess
+            fx = x + wd + 10
+            fw = 26 + len(fixed) * 9
+            b.append(rect(fx, y2, fw, 38, 10, "acc-tint"))
+            b.append(text(fx + fw / 2, y2 + 25, fixed, 16, 600, "acc-t", "middle"))
+            xs[i] = (fx, fw)
+        b.append(f'<line x1="{x + wd / 2}" y1="{y1 + 40}" x2="{x + wd / 2}" y2="{y2 - 4}" class="line" stroke-width="1.4"/>')
+    # the result
+    rx = xs[-1][0] + xs[-1][1] + 40
+    b.append(arrow(rx, y2 + 19, rx + 30, y2 + 19))
+    b.append(rect(rx + 42, y2 - 6, W - 48 - (rx + 42), 50, 12, "acc"))
+    b.append(icon("i-bolt", rx + 54, y2 + 6, 26, "on-acc-icon"))
+    b.append(text(rx + 90, y2 + 16, "4 words in one step", 15, 700, "on-acc"))
+    b.append(text(rx + 90, y2 + 34, "3 guesses kept + its own word", 12, 500, "on-acc"))
+    return svg(W, H, "".join(b).replace('class="on-acc-icon"', 'class="icon" style="color:#ffffff"'),
+               "Speculative decoding: a small helper guesses words, the big model checks them at once")
+
+
+if __name__ == "__main__":
+    (OUT / "how-it-works.svg").write_text(how_it_works(), encoding="utf-8")
+    (OUT / "guess-and-check.svg").write_text(guess_and_check(), encoding="utf-8")
+    for f in ("how-it-works.svg", "guess-and-check.svg"):
+        print(f, (OUT / f).stat().st_size, "bytes")

include/strata/artifact/gguf_reader.hpp

@@ -454,8 +454,9 @@ private:
 // ---- architecture guard (P1.S2). The engine is specialised to ONE model; anything else must be
 // refused with a precise error rather than silently mis-run.
 struct Qwen4ExpGuard {
-    uint32_t block_count = 48, hidden = 2560, experts = 512, experts_used = 10, head_count = 24,
-             head_count_kv = 2;
+    uint32_t block_count = 48, hidden = 2560, experts = 0, experts_used = 0, head_count = 24,
+             head_count_kv = 2;   // 0 = presence-only: pruned variants (GSQ-RCO Coder) legitimately ship
+                                  // fewer experts than the canonical 512; the graph reads the true value
 };
 
 inline std::string check_architecture(const GgufFile& g, const Qwen4ExpGuard& want = {}) {
@@ -477,7 +478,7 @@ inline std::string check_architecture(const GgufFile& g, const Qwen4ExpGuard& wa
     for (const auto& r : reqs) {
         const MetaValue* v = g.get(r.key);
         if (!v) return std::string("missing ") + r.key;
-        if (v->u != r.want)
+        if (r.want && v->u != r.want)
             return std::string(r.key) + " = " + std::to_string(v->u) + ", expected " + std::to_string(r.want);
     }
     return {}; // empty == ok

include/strata/core/expert_source.hpp

@@ -31,6 +31,8 @@
 
 namespace strata::core {
 
+class RemoteExperts;
+
 /// Where one routed expert's bytes come from.
 ///
 /// Phase 2 has NO cache (`phase-2-correct-engine.md`: hit rate `h = 0`), so the only implementation is a
@@ -92,6 +94,8 @@ struct GpuPlanSink {
 struct ExpertDispatch {
     strata::kernels::cpu::ExpertPool* pool = nullptr;
     ExpertSource* src = nullptr;
+    RemoteExperts* remote[3] = {}; ///< optional CUDA1..3 tiers for otherwise CPU-served rows
+    int remote_count = 0;
     int64_t n_expert = strata::kernels::cpu::NE;
 
     /// Counters, for the driver to report rather than for control flow.
@@ -320,7 +324,8 @@ public:
 
     /// Allocates and loads `<pack_dir>/experts.bin`.  Prints nothing; the caller reports `note()` and the load
     /// rate, because those are the two numbers that say whether the arena is the one that was asked for.
-    bool open(const std::string& pack_dir, int64_t n_layers, int64_t n_expert, int threads, std::string& err);
+    bool open(const std::string& pack_dir, int64_t n_layers, int64_t n_expert, int threads, std::string& err,
+              uint64_t max_pinned_bytes = 0);
     /// Plan v0.3 P6: a native pack without experts.bin takes its experts from the model's shard 1.
     void set_gguf(const std::string& shard1) { gguf_ = shard1; }
     void close();

include/strata/core/layer.hpp

@@ -36,6 +36,7 @@
 #include "strata/core/weights.hpp"
 
 #include "strata/kernels/gr.hpp"
+#include "strata/kernels/kv_stream.hpp"
 #include "strata/kernels/ngram.hpp"
 #include "strata/kernels/ple.hpp"
 
@@ -206,10 +207,24 @@ struct QsaState {
     int8_t* v_q = nullptr;
     uint16_t* k_scale = nullptr;
     uint16_t* v_scale = nullptr;
-    int32_t* page_table = nullptr;   ///< (n_pages,) logical page -> physical page
+    /// PR #21: Q4_0 KV with Walsh-Hadamard rotation (qsa_set_kv_q4, kv_q4.hpp): 144 B per cell and head
+    bool kv_q4 = false;
+    uint8_t* k_q4 = nullptr;
+    uint8_t* v_q4 = nullptr;
+    int32_t* page_table = nullptr;   ///< (n_pages,) logical page -> physical page (-1: not resident, streamed)
     int64_t n_pages = 0;
     int64_t max_cells = 0;
 
+    /// KV STREAMING (docs/kv-streaming-design.md, `kv_stream.hpp`). `kv_mode` 0: every page in VRAM, identity
+    /// table (n_slots == n_pages, no host copy). 1: streamed - the authoritative K/V in `host`, `n_slots` pages
+    /// in VRAM, `map` resolves the selection's blocks on device. 2: a ring (the MTP drafter's window) - `n_slots`
+    /// pages at `block % n_slots`, with the host copy for a resume. The pool pointers above point at the slots.
+    int kv_mode = 0;
+    int64_t n_slots = 0;
+    strata::kernels::KvHostPools host;
+    strata::kernels::KvStreamMap map;
+    int64_t idx_pooled_rows = 0;     ///< rows of `idx_pooled` (a ring, which has no indexer, keeps 2)
+
     float* idx_tail = nullptr;       ///< (idx_block - 1, idx_dim): the raw tail of the block being filled
     float* idx_dead = nullptr;       ///< (idx_dim,): the spare slot's key, CONSTANT for the sequence
     float* idx_pooled = nullptr;     ///< (max_cells/idx_block + 2, idx_dim)
@@ -242,13 +257,35 @@ struct QsaState {
 /// Plan v0.3 P7: the RoPE cos/sin table (max_cells x n_rot/2 x 2 floats, 64 MiB at 262K) is identical in every
 /// QSA layer. `with_rope = false` sizes a state that borrows it; `share_rope` points `st` at another state's table
 /// instead of building a copy (the session builds it once, in the first QSA layer).
-uint64_t qsa_state_bytes(const ModelGeometry& g, int64_t max_cells, bool with_rope = true);
+uint64_t qsa_state_bytes(const ModelGeometry& g, int64_t max_cells, bool with_rope = true, int64_t ring_cells = 0);
+/// KV streaming: keep `cells` cells of each QSA layer in VRAM and the rest in pinned host memory (0: all in VRAM,
+/// the default). Set before sizing and initializing the session; a context that fits in `cells` is not streamed.
+/// Also puts the MTP drafter's K/V in a ring of its window (`ring_cells` of qsa_state_bytes/init; -1 forces a fully
+/// resident state).
+void qsa_set_kv_resident(int64_t cells);
+int64_t qsa_kv_resident();
+/// The fewest resident cells a streamed layer may have: one verify window's selections (8 queries x 2,051 cells
+/// in whole blocks) must fit at once, with room to spare.
+int64_t qsa_kv_resident_min();
+/// Pinned host bytes the streamed states hold (their host copies).
+uint64_t qsa_kv_host_bytes();
 /// Plan v0.3 P7: store K/V as INT8 with FP16 scales per 64 values (half the VRAM of FP16). Set before sizing and
 /// initializing the session; default off until gate G-C accepts it.
 void qsa_set_kv_int8(bool enabled);
 bool qsa_kv_int8();
+/// PR #21: store K/V as Q4_0 after a Hadamard rotation (`--kv q4_0`): 576 B per cell, vs 1,056 in INT8.
+void qsa_set_kv_q4(bool enabled);
+bool qsa_kv_q4();
+/// The state's KV format for the block-moving functions of kv_stream.hpp (kKvF16 / kKvInt8 / kKvQ4).
+inline int qsa_kv_format(const QsaState& st) {
+    return st.kv_q4 ? strata::kernels::kKvQ4 : st.kv_int8 ? strata::kernels::kKvInt8 : strata::kernels::kKvF16;
+}
 uint64_t qsa_state_init(const ModelGeometry& g, int64_t max_cells, void* base, QsaState& st,
-                        const QsaState* share_rope = nullptr);
+                        const QsaState* share_rope = nullptr, int64_t ring_cells = 0);
+/// KV streaming: the pools a reader sees (the VRAM slots) and, when streamed, make the selection's blocks resident.
+strata::kernels::QsaAttnPools qsa_attn_pools(const QsaState& st);
+void qsa_kv_resolve(const QsaState& st, const ModelGeometry& g, const int32_t* ids, const int32_t* steps, int64_t n_q,
+                    int64_t cap, void* stream);
 /// Zeroes the pools AND the indexer, so a fresh sequence matches the reference's own `zeros()`.  The KV pool
 /// matters even for cells that are never attended, because `kv_gather` reads whatever the selection names.
 void qsa_state_zero(const QsaState& st, const ModelGeometry& g, void* stream);

include/strata/core/mtp.hpp

@@ -45,7 +45,14 @@ public:
               int64_t window = 32768);
     /// The prompt's length: prefill() skips the cells the attention window can never reach again.
     void set_prompt_len(int64_t n) { prompt_len_ = n; }
+    /// At most this many drafts per round (below max_t - 1): a window longer than the MTP's comes from elsewhere.
+    void set_max_drafts(int k) { max_drafts_ = k; }
     uint64_t vram_bytes() const { return vram_; }
+    /// The draft layer's K/V state (read-only: --serve's STRATA_STATE_HASH check hashes it)
+    const QsaState& kv_state() const { return st_; }
+    /// KV streaming: refill the ring of the drafter's window from its host copy for a sequence that continues at
+    /// `upto` (a conversation-cache resume). No-op unless the drafter's K/V is a ring.
+    void kv_restore(int64_t upto);
     /// The main model's embedding and head, and the verify window's final residuals (T rows, hc*n_embd each).
     bool bind(const WeightTable& wt, const NativeHead* head, const float* window_R, std::string& err);
 
@@ -82,6 +89,8 @@ private:
     const NativeHead* head_ = nullptr;
     const float* window_R_ = nullptr;
     int max_t_ = 0;
+    int device_ = -1;   ///< the device `load` ran on: the public calls switch to it (layer split)
+    int max_drafts_ = 1 << 30;
     int64_t n_vocab_ = 0;
     uint64_t vram_ = 0;
     cudaStream_t cs_ = nullptr;

include/strata/core/on_device.hpp

@@ -0,0 +1,22 @@
+#pragma once
+// Multi-GPU: make `device` current for a scope and restore the caller's device after it.  An object that owns CUDA
+// streams, graphs and buffers on one device (a verify stage, the drafter) wraps its public calls in this, so the
+// caller's thread may be on any device.  A negative device, or the device already current, does nothing.
+
+#include <cuda_runtime.h>
+
+namespace strata::core {
+
+struct OnDevice {
+    int previous = -1;
+    explicit OnDevice(int device) {
+        int cur = 0;
+        if (device >= 0 && cudaGetDevice(&cur) == cudaSuccess && cur != device && cudaSetDevice(device) == cudaSuccess)
+            previous = cur;
+    }
+    ~OnDevice() { if (previous >= 0) cudaSetDevice(previous); }
+    OnDevice(const OnDevice&) = delete;
+    OnDevice& operator=(const OnDevice&) = delete;
+};
+
+}  // namespace strata::core

include/strata/core/pinned.hpp

@@ -39,7 +39,8 @@ struct PinnedArena {
     explicit PinnedArena(uint64_t bytes, uint64_t slice = 0);
     /// Plan v0.3 P6: slices of different sizes (one per layer of a native pack), given as their start offsets
     /// followed by the end of the last one.  `slice_starts` holds the registered ones.
-    PinnedArena(uint64_t bytes, const std::vector<uint64_t>& bounds);
+    /// `max_pinned_bytes`: optional cap on CUDA registration. 0 preserves the normal unrestricted path.
+    PinnedArena(uint64_t bytes, const std::vector<uint64_t>& bounds, uint64_t max_pinned_bytes = 0);
     std::vector<uint64_t> slice_starts;
     ~PinnedArena();
     PinnedArena(const PinnedArena&) = delete;

include/strata/core/progress.hpp

@@ -0,0 +1,51 @@
+// include/strata/core/progress.hpp - issue #29: whether a request is still moving, and where it is.
+//
+// The engine's host loop and the GPU wait on each other through flags; a protocol bug there does not crash, it
+// spins forever (the GPU "100%", one CPU core busy, no tokens).  `--serve` runs a watchdog thread over this: a
+// request whose heartbeat stops for `STRATA_WATCHDOG_S` seconds (default 60; 0 = off) ends the engine with the
+// stage it was stuck in, and the server starts it again instead of hanging.  Tokens, prompt chunks and verify
+// windows beat; the stage is two relaxed stores per layer, which the token path does not notice.
+#pragma once
+
+#include <atomic>
+#include <chrono>
+#include <cstdint>
+#include <cstdio>
+
+namespace strata::core {
+
+struct Progress {
+    std::atomic<uint64_t> beats{0};
+    std::atomic<bool> busy{false};
+    std::atomic<const char*> where{"idle"};
+    std::atomic<int64_t> detail{-1};
+    std::atomic<int64_t> since_ms{0};    ///< when `where` was set (steady clock): how long a stage has lasted
+    std::atomic<uint64_t> ticks{0};      ///< layers served: a window that still moves, slowly, against one that stopped
+};
+
+inline int64_t progress_now_ms() {
+    return std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now().time_since_epoch())
+        .count();
+}
+
+/// What the watchdog prints about a part of the engine when it fires (the expert pool, the verify window).
+using DiagFn = void (*)(std::FILE*);
+inline std::atomic<DiagFn>& diag_pool_fn() { static std::atomic<DiagFn> f{nullptr}; return f; }
+inline std::atomic<DiagFn>& diag_verify_fn() { static std::atomic<DiagFn> f{nullptr}; return f; }
+
+inline Progress& progress() {
+    static Progress p;
+    return p;
+}
+
+inline void progress_at(const char* where, int64_t detail = -1) {
+    Progress& p = progress();
+    p.where.store(where, std::memory_order_relaxed);
+    p.detail.store(detail, std::memory_order_relaxed);
+    p.since_ms.store(progress_now_ms(), std::memory_order_relaxed);
+}
+
+inline void progress_beat() { progress().beats.fetch_add(1, std::memory_order_relaxed); }
+inline void progress_tick() { progress().ticks.fetch_add(1, std::memory_order_relaxed); }
+
+}  // namespace strata::core

include/strata/core/remote_experts.hpp

@@ -0,0 +1,85 @@
+#pragma once
+
+#include "strata/core/expert_cache.hpp"
+#include "strata/core/expert_source.hpp"
+
+#include <cuda_runtime.h>
+
+#include <cstdint>
+#include <string>
+#include <utility>
+#include <vector>
+
+namespace strata::core {
+
+/// A static, profile-filled expert tier on another CUDA device. CUDA0 keeps all
+/// dense weights and state; results return through the existing pinned CPU rows.
+class RemoteExperts {
+public:
+    RemoteExperts() = default;
+    ~RemoteExperts();
+    RemoteExperts(const RemoteExperts&) = delete;
+    RemoteExperts& operator=(const RemoteExperts&) = delete;
+
+    /// Initialise the device before the host expert arena registers
+    /// tens of GiB of portable mapped memory with CUDA.
+    static bool preflight(int device, double& free_gib, std::string& err);
+    bool open(int device, int slots, int64_t layers, int64_t experts,
+              const std::vector<std::pair<int32_t, int32_t>>& ranked,
+              const ExpertCache& primary, ExpertSource& source,
+              std::vector<uint8_t>& claimed, std::string& err);
+    void close();
+
+    /// `kind` is the primary verifier's classification (-1 = CPU candidate),
+    /// or null on the one-token path. Entries already served on CUDA0 are excluded.
+    bool begin(int64_t layer, const float* x, const int32_t* ids, int64_t n_tok,
+               int64_t k, const int32_t* kind, const int32_t* primary_res,
+               std::string& err);
+    bool owns(int64_t index) const { return owned_[(size_t) index] != 0; }
+    bool finish(float* out, std::string& err);
+    int64_t resident() const { return cache_.resident(); }
+    int64_t computed() const { return computed_; }
+    int64_t launched_layers() const { return launched_layers_; }
+    double gib() const { return cache_.gib(); }
+    uint64_t returned_bytes() const { return returned_bytes_; }
+    uint64_t full_row_bytes() const { return full_row_bytes_; }
+    /// host time spent in begin() (staging + launches) and in finish() (waiting for this GPU), cumulative
+    double ms_begin() const { return ms_begin_; }
+    double ms_wait() const { return ms_wait_; }
+
+private:
+    int device_ = -1;
+    int64_t n_expert_ = 0;
+    int32_t groups_ = 0;
+    int64_t computed_ = 0;
+    int64_t launched_layers_ = 0;
+    uint64_t returned_bytes_ = 0;
+    uint64_t full_row_bytes_ = 0;
+    double ms_begin_ = 0, ms_wait_ = 0;
+    ExpertCache cache_;
+    cudaStream_t stream_ = nullptr;
+    float* h_x_ = nullptr;
+    float* h_out_ = nullptr;
+    void* h_meta_ = nullptr;
+    float* d_x_ = nullptr;
+    float* z_x_ = nullptr;     ///< h_x_ as the helper GPU sees it (zero-copy: no input copy per layer)
+    float* z_out_ = nullptr;   ///< h_out_ as the helper GPU sees it (zero-copy: no result copy)
+    bool zero_copy_ = false;
+    float* d_out_ = nullptr;
+    uint8_t* d_q8_ = nullptr;
+    float* d_scales_ = nullptr;
+    void* d_scratch_ = nullptr;
+    void* d_meta_ = nullptr;  ///< one contiguous upload of grouped indices, instead of five small copies
+    int32_t* d_start_ = nullptr;
+    int32_t* d_dst_ = nullptr;
+    int32_t* d_tok_ = nullptr;
+    int32_t* d_count_ = nullptr;
+    unsigned long long* d_ptr_ = nullptr;
+    std::vector<uint8_t> owned_;
+    std::vector<uint8_t> layers_present_;
+    std::vector<int32_t> group_of_, group_id_;
+    std::vector<int32_t> start_, dst_, tok_, original_row_;
+    std::vector<unsigned long long> ptr_;
+};
+
+} // namespace strata::core

include/strata/core/verify.hpp

@@ -21,9 +21,12 @@
 // selection, native indexer) and a profile-filled VRAM expert tier with its residency table on the device.
 #pragma once
 
+#include <cstdio>
+
 #include "strata/core/expert_source.hpp"
 #include "strata/core/layer.hpp"
 #include "strata/core/session.hpp"
+#include "strata/kernels/sampler.hpp"
 
 #include <cuda_runtime.h>
 
@@ -51,6 +54,9 @@ public:
     Verifier(const Verifier&) = delete;
     Verifier& operator=(const Verifier&) = delete;
 
+    /// The watchdog's view of the window in flight (issue #31): the layer, the GPU's sequence, the flags.
+    void diag(std::FILE* f) const;
+
     /// `max_t` <= kVerifyMaxT.  `head` may be null (the canonical head is then run per token).
     bool init(const WeightTable& wt, const ModelGeometry& g, SessionState& ss, const VerifyHits& hits,
               const NativeHead* head, int max_t, std::string& err);
@@ -58,13 +64,52 @@ public:
     /// One window: `tokens[0..T)` at positions pos0.., the pool served per layer; `out[t]` = argmax after token t.
     /// The PLE rows are gathered here from `ss.ple_prev` and the tokens.  Captures the T-token graph on first use.
     bool run(int T, const int32_t* tokens, int64_t pos0, PoolMultiFn pool, void* user, int32_t* out, std::string& err);
+    /// The sampling the verify window's head applies (temperature / top_p / top_k / seed).  Set per
+    /// request; greedy by default.  The sampling itself runs OUTSIDE the captured graph - its
+    /// parameters would otherwise be baked forever - so this can change between requests freely.
+    void set_sampling(const strata::kernels::SamplerParams& sp) {
+        sampling_ = sp;   // row t of a window at pos0 draws Philox(seed, pos0 + t): see run()
+        if (next_) next_->set_sampling(sp);
+    }
+
+    /// The penalty histories for `sampling_.penalty_last_n`: ONE ROW PER WINDOW ROW, T rows of `history_len`
+    /// int32 slots at that stride (`strata::kernels::penalty_rows` builds them), most recent token LAST, unused
+    /// front slots -1 (the kernel reads only the tail window).  Row t follows the window's drafts 1..t - staging
+    /// row 0 alone (before 0.1.19) left the drafted rows with unwritten histories.  Null disables the penalties
+    /// entirely - the neutral run's sampling call is byte-for-byte what it was.  The engine re-uploads the rows
+    /// before every window; the buffer must hold kVerifyMaxT rows and stay alive across the request.
+    void set_history(const int32_t* history, int history_len) {
+        hist_d_ = history;
+        hist_len_ = history_len;
+        if (next_) next_->set_history(history, history_len);
+    }
+    /// Off: `run` skips the request's head sampling and `out` is the recorded greedy pick.  For windows whose
+    /// picks are discarded - a prompt read through windows commits every token - so they cost no sampler launch
+    /// or sync and never read a history staged for another position.
+    void set_head_sampling(bool on) { head_sampling_ = on; if (next_) next_->set_head_sampling(on); }
+
+    /// LAYER SPLIT (multi-GPU): this verifier runs layers [layer_begin, layer_end) of every window.  A stage that
+    /// does not start at layer 0 takes its residual from `handoff_in` instead of embedding the tokens; a stage that
+    /// does not end at the last layer writes its residual to `handoff_out` and has no head.  The hand-off holds,
+    /// per token, the residual R (hc x n_embd), the last layer's pending write bo (n_embd) and inject (hc): the next
+    /// stage folds that write into its first read exactly as the unsplit window does, so the split is bit-exact.
+    /// Both pointers must be device-visible (mapped pinned memory, portable when the stages are on two devices).
+    /// Set before `init`.  Default: the whole model, no hand-off.
+    void set_stage(int64_t layer_begin, int64_t layer_end, const float* handoff_in, float* handoff_out) {
+        lb_ = layer_begin; le_ = layer_end; hand_in_ = handoff_in; hand_out_ = handoff_out;
+    }
+    /// The next stage: `run` and `commit` continue into it (its pool calls get `next_user`); sampling settings
+    /// and `final_R` are the last stage's.
+    void set_next(Verifier* next, void* next_user) { next_ = next; next_user_ = next_user; }
+    /// floats per token in a hand-off buffer
+    static int64_t handoff_floats(const ModelGeometry& g) { return (int64_t) g.hc * g.n_embd + g.n_embd + g.hc; }
 
     /// Keep the first `n_keep` (1..T) tokens of the last window; advances `ss.ple_prev` by them.
     bool commit(int n_keep, std::string& err);
 
     /// Token t's residual after the last layer, (hc, n_embd) on the device, valid until the next `run`.
     const float* final_R(int t) const;
-    const float* final_R_all() const { return R_; }
+    const float* final_R_all() const { return next_ ? next_->final_R_all() : R_; }
 
     /// The GPU plan the pool writes each layer (VRAM hits + the PCIe share of the misses); give it to the
     /// dispatch (`ExpertDispatch::plan`) before the first `run`.
@@ -83,6 +128,22 @@ public:
 
 private:
     bool capture(int T, std::string& err);
+    strata::kernels::SamplerParams sampling_ = [] {
+        strata::kernels::SamplerParams s;
+        s.greedy = true;
+        s.temperature = 0.0f;
+        return s;
+    }();   ///< greedy by default; per-request via set_sampling
+    const int32_t* hist_d_ = nullptr;   ///< penalty-history row (set_history); null = no penalties apply
+    int hist_len_ = 0;
+    bool head_sampling_ = true;          ///< set_head_sampling
+    int device_ = -1;                    ///< the device `init` ran on: run/commit switch to it (layer split)
+    int64_t lb_ = 0, le_ = -1;           ///< set_stage: the layers this verifier runs (-1: to the last)
+    const float* hand_in_ = nullptr;
+    float* hand_out_ = nullptr;
+    Verifier* next_ = nullptr;
+    void* next_user_ = nullptr;
+    bool ple_stage() const { return lb_ <= 1 && 1 < le_; }   ///< holds layer 1, where the PLE block runs
     bool capture_commit(std::string& err);
     bool record_window(int T, cudaStream_t cs, std::string& err);
 

include/strata/kernels/cpu/iq_avx2.hpp

@@ -0,0 +1,26 @@
+// include/strata/kernels/cpu/iq_avx2.hpp - AVX-2 multi-token dot products for the i-quant expert
+// formats (IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S) against Q8_K activations (ggml's block_q8_K).
+//
+// The AVX-512 kernel's scheme on the CPUs without it (AMD Zen 2/3, Intel Core 12th-14th gen): the weights of a
+// 32-value chunk are decoded once per verify window and every token applies them with five instructions.
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+namespace strata::kernels::cpu {
+
+bool iq256_supported(int ggml_type) noexcept;
+/// ff[t][r] = silu(gate_r . a[t]) * (up_r . a[t]), rows [r0, r1); gate rows at blob, up rows at blob + up_off.
+void iq256_gu_rows(int ggml_type, const uint8_t* blob, size_t gu_row, size_t up_off, int n, const void* const* act,
+                   int nt, float* const* ff, int r0, int r1);
+/// out[t][r] = w_r . a[t], rows [r0, r1).
+void iq256_rows(int ggml_type, const uint8_t* w, size_t row_bytes, int n, const void* const* act, int nt,
+                float* const* out, int r0, int r1);
+
+/// out[t][r] = w_r . h[t] for IQ4_NL (type 20) rows against Q8_0 activations (ggml's block_q8_0).
+/// IQ4_NL is a 32-value-block format, so this does not go through iq256_rows (QK_K blocks, Q8_K acts).
+void iq4nl256_down_rows(const uint8_t* w, size_t row_bytes, int n, const void* const* hq, int nt,
+                        float* const* out, int r0, int r1);
+
+}  // namespace strata::kernels::cpu

include/strata/kernels/cpu/pool.hpp

@@ -14,16 +14,29 @@
 // THE COMPLETION PROTOCOL, because this is where a pool usually goes wrong.  `run()` waits for `done == n`
 // AND for every worker to PARK.  Waiting only for `done` is not enough: a worker can still be inside the
 // drain loop after its last `done` increment, and the host resetting `head` underneath it would let that
-// worker claim a job from the NEXT batch before the next batch has been published.  The
-// `done`-then-`parked` pair makes the handover unambiguous, and the second wait costs a few hundred cycles
-// against a layer that takes milliseconds.
+// worker claim a job from the NEXT batch before the next batch has been published.
+//
+// **AND `parked` IS NOT ENOUGH EITHER (issue #29).**  A worker that went to sleep (after `kSpinBeforeSleep`) is
+// still counted as parked when it wakes, so for a moment after it has seen a new epoch the host believes it is
+// idle.  On a card with most experts in VRAM the workers sleep in the middle of a request, tiny batches finish
+// before a sleeper is awake, and that moment comes round constantly: a late worker could claim from the NEXT
+// batch while the host was still writing it, run a job twice, or add to `done` after the host had reset it -
+// and `done != n` then never ended, with the GPU waiting on the pool forever.  So every claim carries its
+// batch: `head` is one 64-bit word `epoch | njobs | index`, a claim is a CAS that only succeeds for the epoch
+// the worker woke for, and a late worker's claim simply fails.  A claim that succeeds belongs to the current
+// batch, whose description the host cannot change until that job's `done` has landed.
 #pragma once
 
 #include "strata/kernels/cpu/expert.hpp"
 #include "strata/kernels/cpu/native_expert.hpp"
 
 #include <atomic>
+#include <cstdio>
+#include <memory>
+#include <chrono>
+#include <condition_variable>
 #include <cstdint>
+#include <mutex>
 #include <thread>
 #include <vector>
 
@@ -89,6 +102,8 @@ public:
     /// threads on six cores. `false` is the A/B arm and exists so the change is measurable rather than
     /// asserted - the counter it moves is `pool phases ... drain`, which is host-side and needs no profiler.
     explicit ExpertPool(int n_workers = 0, bool pin = true, bool host_works = true);
+    /// The watchdog's view of the pool (issue #31): the batch, the counters, every thread's state.
+    void diag(std::FILE* f) const;
     ~ExpertPool();
     ExpertPool(const ExpertPool&) = delete;
     ExpertPool& operator=(const ExpertPool&) = delete;
@@ -136,10 +151,31 @@ public:
         repark = ms_repark_;
     }
 
+    /// **A PARKED WORKER SPINS FOR THIS LONG, THEN SLEEPS.**  The park is a `_mm_pause` spin because a layer's
+    /// batches are microseconds apart and a wake-up from the OS costs more than that.  But a spin that never ends
+    /// keeps every worker's core at 100% while the engine waits for a request - issue #4, "CPU 50% even when
+    /// doing nothing" (7 of the 5700X's 16 threads).  Between requests the workers block on `sleep_cv_`.  NOT only
+    /// between requests: with most experts in VRAM (a 24 GB card) many layers have no CPU work, so the workers
+    /// also sleep mid-request - which the claim protocol above must survive (issue #29).
+    /// `STRATA_POOL_SPIN_US` overrides it (a test knob: a short spin makes the workers sleep constantly).
+    static constexpr std::chrono::milliseconds kSpinBeforeSleep{20};
+    /// A pool wait that sees no completion for this long is a bug; the engine stops with a message instead of
+    /// spinning forever, and the server starts it again (issue #29).
+    static constexpr std::chrono::seconds kStall{60};
+
 private:
     void worker(int id);
-    void drain(int id, ExpertScratch& scratch);
+    void drain(int id, ExpertScratch& scratch, uint32_t epoch);
     void run_phase(int mode, int n_tasks);
+    /// Claim the next job of batch `epoch`, or -1 (that batch is exhausted, or it is not the current one).
+    int claim(uint32_t epoch);
+    /// Publish the batch whose description the caller has just written: reset `done`, then `head`, then the epoch.
+    uint32_t begin_batch(int n);
+    /// The host's waits, bounded by `kStall`.
+    void wait_parked(const char* what);
+    void wait_done(int n);
+    /// Bump `epoch_`, and wake the workers that went to sleep.  Every publish goes through here.
+    void publish();
 
     int n_ = 0;
     bool host_works_ = true;
@@ -164,11 +200,24 @@ private:
     //
     // The counter was diagnostic only - `pauses()` was read in one place, to print a number nothing branched on
     // - so it is deleted rather than amortised.  `alignas(64)` then stops the remaining four sharing.
-    alignas(64) std::atomic<uint32_t> head_{0};
+    // issue #31 diagnostics: each worker's state (kParked, kSleeping, kBetween, or the job it runs) and the host's
+    // (kIdle, kWaitParked, kWaitDone, or its job), printed by the serve watchdog through `diag`
+    static constexpr int32_t kParked = -1, kSleeping = -2, kBetween = -3, kIdle = -10, kWaitParked = -11,
+                             kWaitDone = -12;
+    std::unique_ptr<std::atomic<int32_t>[]> wstate_;
+    std::atomic<int32_t> hstate_{kIdle};
+    std::atomic<int64_t> hstate_ms_{0};
+    alignas(64) std::atomic<uint64_t> head_{0};   // epoch << 32 | njobs << 16 | next index (issue #29)
     alignas(64) std::atomic<uint32_t> done_{0};
     alignas(64) std::atomic<uint32_t> parked_{0};
     alignas(64) std::atomic<uint32_t> epoch_{0};
     alignas(64) std::atomic<bool> stop_{false};
+    // The sleep after `kSpinBeforeSleep`.  `sleepers_` is how `publish` knows whether anyone needs waking, so the
+    // token path pays one uncontended load per publish and never takes the mutex while the workers spin.
+    alignas(64) std::atomic<uint32_t> sleepers_{0};
+    std::mutex sleep_mu_;
+    std::condition_variable sleep_cv_;
+    std::chrono::microseconds spin_before_sleep_{kSpinBeforeSleep};
     std::vector<std::thread> threads_;
     std::vector<ExpertScratch> scratch_;   // one per worker: no allocation, no false sharing of the hot data
     // run_split state: mode 0 = whole experts, 1 = gate/up row parts, 2 = down row parts

include/strata/kernels/cvec.hpp

@@ -0,0 +1,56 @@
+// include/strata/kernels/cvec.hpp - a control vector on the residual stream: the engine's side of llama.cpp's
+// `llama_adapter_cvec` with the `--cvec-mode` / `--cvec-dir` extension of the `experimental-speed-projection`
+// package (its `02-cvec-projection-mode.patch`, and `01-qwen4exp-cvec-hooks.patch` for WHERE it applies).
+//
+// After layer l's FFN write - llama.cpp's `build_cvec(res_hc, il)` right after the layer's second
+// `build_hc_combine` - every one of the hc residual streams h of every token becomes
+//
+//   project:  h <- h - s_l (h . v_l) v_l        v_l = d / |d|, s_l = |d|, d = the scaled direction
+//   add:      h <- h + d                        llama.cpp's stock additive control vector
+//
+// for l in [first, last].  `d` is the GGUF's `direction.<src>` times the file's scale (summed over files), with
+// src = l (per-layer) or one fixed layer (single:L, project mode only), exactly as llama.cpp's loader places it;
+// layer 0 has no direction.  Off unless a vector is loaded (--control-vector-scaled); a loaded one is switched
+// per request through a device flag, so the captured graphs are the same either way.  With the flag off the
+// result is bit-identical to the engine without a vector: the kernel then only applies the pending residual
+// write, with the fused read's own arithmetic.
+#pragma once
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+namespace strata::kernels {
+
+struct Cvec {
+    const float* dir = nullptr;   ///< device, n_layers * n_embd: project = the unit v_l, add = d_l; zero rows elsewhere
+    const float* s = nullptr;     ///< device, n_layers: project = s_l, add = 1; 0 = the layer is not steered
+    const int* on = nullptr;      ///< device int: 0 = this request runs the stock model
+    int mode = 0;                 ///< 0 = project, 1 = add
+    int first = 0, last = -1;     ///< the steered layers, inclusive
+    int64_t n_embd = 0, hc = 0;
+    bool loaded() const { return dir != nullptr; }
+    /// Layer l's FFN write is followed by the vector (with a real direction there).  Decided at load time, never
+    /// per request: it changes where the residual write happens, so the graphs depend on it.
+    bool covers(int64_t l) const { return loaded() && l >= first && l <= last && l < (int64_t) steered.size() && steered[(size_t) l]; }
+    std::vector<bool> steered;    ///< host copy of s_l != 0
+};
+
+/// The loaded vector (empty until `cvec_upload`).
+const Cvec& cvec();
+
+/// Upload a vector built by the loader: `dir` is n_layers * n_embd, `s` n_layers (see `Cvec`).  Starts ON.
+bool cvec_upload(const std::vector<float>& dir, const std::vector<float>& s, int mode, int first, int last,
+                 int64_t n_embd, int64_t hc, std::string& err);
+
+/// The per-request switch.  Synchronizes the device when it changes, so call it between requests.
+void cvec_set_enabled(bool on);
+bool cvec_enabled();
+
+/// Layer `layer`'s vector on T tokens' residual stacks (`R + t * r_ld`, hc streams of n_embd).  With `write`, the
+/// pending FFN write `R += bo * 2 sigmoid(inj / hc)` (the fused read's arithmetic) is applied first, for callers
+/// whose writes are folded into the next layer's read; without it, R must already hold the layer's output.
+void cvec_apply(float* R, int64_t layer, int64_t T, int64_t r_ld, const float* bo, int64_t bo_ld, const float* inj,
+                int64_t inj_ld, bool write, void* stream);
+
+}  // namespace strata::kernels

include/strata/kernels/kv_q4.hpp

@@ -0,0 +1,61 @@
+// include/strata/kernels/kv_q4.hpp - Q4_0 KV storage with Walsh-Hadamard rotation for the QSA layers (`--kv q4_0`).
+//
+// From PR #21 (code-martin). K and V are rotated by the orthonormal 256-point Hadamard matrix H before they are
+// quantized to ggml's q4_0 (32 values per block: an fp16 scale + 16 bytes of 4-bit codes, 144 B per head per cell,
+// 576 B per cell for K and V of both heads, vs 1,056 B in INT8). The rotation spreads a head's outlier channels
+// over all 256 dimensions, which is what makes 4 bits usable. The query is rotated the same way, so
+// <Hq, Hk> = <q, k> and the scores are unchanged, and the attention output (a mix of rotated values) is rotated
+// back with H (self-inverse). What is lossy is the 4-bit rounding itself: measured, not assumed, in
+// bench/results/2026-09-27-kv-q4.
+//
+// With KV streaming (--kv-resident) the codes live in the host copy and the resident slots like INT8's (KvHostPools
+// k_q4/v_q4, the same block granule), so both options combine.
+#pragma once
+
+#include "strata/kernels/qsa.hpp"
+
+#include <cstdint>
+
+namespace strata::kernels {
+
+inline constexpr int QK4_0 = 32;
+
+#pragma pack(push, 1)
+struct block_q4_0 {
+    uint16_t d;             // scale (fp16 bits)
+    uint8_t qs[QK4_0 / 2];  // 32 4-bit codes: element j in the low nibble of qs[j], j + 16 in the high one
+};
+#pragma pack(pop)
+
+static_assert(sizeof(block_q4_0) == 18, "block_q4_0 must be 18 bytes");
+
+/// Bytes per cell and KV head: 8 blocks of 32 for head_dim 256 -> 144.
+inline uint64_t kv_q4_bytes_per_head(int head_dim) { return (uint64_t) (head_dim / QK4_0) * sizeof(block_q4_0); }
+
+/// Bytes per cell (one token, one layer): K and V of every KV head.
+inline uint64_t kv_q4_bytes_per_cell(const QsaShapes& s) {
+    return (uint64_t) s.n_head_kv * kv_q4_bytes_per_head((int) s.head_dim) * 2;
+}
+
+/// Orthonormal Fast Walsh-Hadamard Transform of rows of 256 floats (scale 1/16): its own inverse.
+void fwht256_cuda(const float* src, float* dst, int64_t n_rows, void* stream);
+inline void fwht256_inplace_cuda(float* data, int64_t n_rows, void* stream) { fwht256_cuda(data, data, n_rows, stream); }
+
+/// Append the (already rotated) cell at step[kStepPos]. With a host copy (KV streaming) it is written there too,
+/// and to VRAM only if its block is resident.
+void kv_append_q4_step(uint8_t* k_q4, uint8_t* v_q4, const int32_t* page_table, const int32_t* step,
+                       const float* kcur, const float* vcur, const QsaShapes& s, void* stream,
+                       const KvHostPools* host = nullptr);
+
+/// The prompt path: T consecutive (rotated) cells from pos0, K/V [T, n_head_kv, 256]; also into `stage` (identity
+/// layout, the one-layer staging pool of a streamed session) when given.
+void kv_append_q4(uint8_t* k_q4, uint8_t* v_q4, const int32_t* page_table, int64_t pos0, int64_t T, const float* K,
+                  const float* V, const QsaShapes& s, void* stream, const KvHostPools* host = nullptr,
+                  const KvHostPools* stage = nullptr);
+
+/// Gather step[kStepWidth] cells named by `ids` into FP16 scratch [id][kv_head][head_dim] (still rotated).
+void kv_gather_q4_step(const uint8_t* k_q4, const uint8_t* v_q4, const int32_t* page_table, const int32_t* ids,
+                       const int32_t* step, int64_t max_ids, const QsaShapes& s, uint16_t* k_scratch,
+                       uint16_t* v_scratch, void* stream);
+
+}  // namespace strata::kernels

include/strata/kernels/kv_q8.hpp

@@ -12,6 +12,7 @@
 // the FP16 pools: codes `[page][kv_head][page_size][head_dim]`, scales `[page][kv_head][page_size][head_dim/64]`.
 #pragma once
 
+#include "strata/kernels/kv_stream.hpp"
 #include "strata/kernels/qsa.hpp"
 
 #include <cstdint>
@@ -25,9 +26,11 @@ inline uint64_t kv_q8_bytes_per_cell(const QsaShapes& s) {
     return (uint64_t) s.n_head_kv * s.head_dim * 2 + (uint64_t) s.n_head_kv * (s.head_dim / KV_Q8_GROUP) * 2 * 2;
 }
 
-/// Append the cell at step[kStepPos] (graph-capturable: position and page come from device memory).
+/// Append the cell at step[kStepPos] (graph-capturable: position and page come from device memory). With a host
+/// copy (KV streaming, `kv_stream.hpp`) the cell is written there too, and to VRAM only if its block is resident.
 void kv_append_q8_step(int8_t* k_q, int8_t* v_q, uint16_t* k_scale, uint16_t* v_scale, const int32_t* page_table,
-                       const int32_t* step, const float* kcur, const float* vcur, const QsaShapes& s, void* stream);
+                       const int32_t* step, const float* kcur, const float* vcur, const QsaShapes& s, void* stream,
+                       const KvHostPools* host = nullptr);
 
 /// Gather step[kStepWidth] cells named by `ids` into FP16 scratch `[id][kv_head][head_dim]`; the grid is sized by
 /// `max_ids` (capacity), the kernel reads the real count from `step`.

include/strata/kernels/kv_stream.hpp

@@ -0,0 +1,98 @@
+// include/strata/kernels/kv_stream.hpp - KV streaming for the QSA layers (docs/kv-streaming-design.md).
+//
+// A streamed QSA layer keeps its AUTHORITATIVE K/V in pinned, device-mapped host memory, laid out exactly as a
+// fully resident pool with the identity page table would be. VRAM holds a pool of `n_slots` pages (one page =
+// `page_size` = 4 cells = one indexer block, all KV heads, K and V: 4,224 B in int8), and the page table becomes a
+// RESIDENCY MAP: `page_table[block]` is the slot holding that block, or -1.
+//
+// Every reader (`qsa_decode_attn_*`, `kv_gather_*`) is unchanged: it goes through the page table as before. What
+// changes is that after the selection and before the attention, `kv_stream_resolve` makes every block the
+// selection names resident - on device, inside the captured graphs:
+//   1. one block of 1,024 threads walks the selected ids: a resident block gets this call's epoch and its clock
+//      reference bit, a missing block is claimed once (page_table -1 -> -2) into a miss list;
+//   2. the same block picks one victim slot per miss with a CLOCK sweep (second chance) that never takes a slot
+//      this call uses, re-points the page table (old block -> -1, new block -> slot);
+//   3. a copy kernel reads the missed blocks from the host copy (zero-copy, over PCIe) into their slots.
+// Writers (`kv_append_*`) write the host copy always, and the slot only if the block is resident, so the slots
+// never go stale and nothing needs invalidating - not after a rejected draft, not on a conversation-cache resume.
+//
+// A RING (the MTP drafter, which only ever reads its last `window` cells) uses the same host copy with a static
+// table `block -> block % n_slots` and no resolve: the ring is refilled from the host copy on a resume.
+#pragma once
+
+#include "strata/kernels/qsa.hpp"
+#include "strata/kernels/qsa_decode_attn.hpp"
+
+#include <cstdint>
+
+namespace strata::kernels {
+
+/// The host copy of one layer's K/V: device-mapped pointers (UVA) into pinned memory, identity layout
+/// `[block][kv_head][page_size][head_dim]`. All null when the layer is fully resident.
+struct KvHostPools {
+    uint16_t* k_pool = nullptr;   ///< fp16 mode
+    uint16_t* v_pool = nullptr;
+    int8_t* k_q = nullptr;        ///< int8 mode: codes
+    int8_t* v_q = nullptr;
+    uint16_t* k_scale = nullptr;  ///< int8 mode: fp16 scale per 64 values
+    uint16_t* v_scale = nullptr;
+    uint8_t* k_q4 = nullptr;      ///< q4_0 mode (kv_q4.hpp): block_q4_0 codes, 144 B per cell and head
+    uint8_t* v_q4 = nullptr;
+    bool present() const { return k_pool != nullptr || k_q != nullptr || k_q4 != nullptr; }
+};
+
+/// The KV storage format, for the functions below that move whole blocks (`fmt`): fp16, int8 (+ scales), q4_0.
+/// (A bool `int8` argument still reads as kKvF16 / kKvInt8.)
+enum KvFormat : int { kKvF16 = 0, kKvInt8 = 1, kKvQ4 = 2 };
+
+/// The residency map of a streamed layer, all device memory at fixed addresses (the graphs bake them in).
+struct KvStreamMap {
+    int32_t* page_table = nullptr;  ///< (n_blocks,) block -> slot, or -1
+    int32_t* slot_block = nullptr;  ///< (n_slots,) slot -> block, or -1
+    int32_t* slot_stamp = nullptr;  ///< (n_slots,) the resolve epoch that last used the slot
+    int32_t* slot_ref = nullptr;    ///< (n_slots,) clock reference bit
+    int32_t* ctl = nullptr;         ///< kKvCtlInts: epoch, hand, misses of the last call, overflow, u64 counters
+    int32_t* miss_block = nullptr;  ///< (n_slots,)
+    int32_t* miss_slot = nullptr;   ///< (n_slots,)
+    int64_t n_blocks = 0;
+    int64_t n_slots = 0;
+};
+
+inline constexpr int kKvCtlInts = 16;   ///< [0] epoch [1] hand [2] misses [3] overflow; u64 at [4] misses, [6] lookups, [8] calls
+/// Bytes of the map's arrays besides the page table (which the state already has): 5 per-slot ints + ctl.
+inline uint64_t kv_stream_map_bytes(int64_t n_slots) { return (uint64_t) n_slots * 4 * 5 + kKvCtlInts * 4; }
+
+/// Every block evicted: page table all -1, slots free, counters zero.
+void kv_stream_reset(const KvStreamMap& m, void* stream);
+
+/// Make every block named by the selections of `n_q` queries resident (ids [n_q][cap], width from
+/// steps[q * kStepCount + kStepWidth]). Capturable. `n_slots` must hold the distinct blocks of one call
+/// (n_q x (cap / page_size + 2)); a call that cannot sets ctl[3] (see `kv_stream_counters`).
+void kv_stream_resolve(const KvStreamMap& m, const QsaAttnPools& slots, const KvHostPools& host, int fmt,
+                       const int32_t* ids, const int32_t* steps, int64_t n_q, int64_t cap, const QsaShapes& s,
+                       void* stream);
+
+/// The static ring table `block -> block % n_slots`.
+void kv_ring_table(int32_t* page_table, int64_t n_blocks, int64_t n_slots, void* stream);
+
+/// Copy blocks [b0, b1) of the host copy into the pool pages `page_table` names (host-side table `phys(b)`
+/// computed as `b % n_slots`: the ring restore). Not capturable.
+void kv_ring_restore(const QsaAttnPools& slots, const KvHostPools& host, int fmt, int64_t b0, int64_t b1,
+                     int64_t n_slots, const QsaShapes& s, void* stream);
+
+/// Copy the first `n_blocks` blocks of the host copy into a fully resident (identity-layout) pool: the prompt
+/// path's staging of one layer. Not capturable (DMA).
+void kv_stage_from_host(const QsaAttnPools& stage, const KvHostPools& host, int fmt, int64_t n_blocks,
+                        const QsaShapes& s, void* stream);
+
+struct KvStreamCounters {
+    uint64_t misses = 0, lookups = 0, calls = 0;
+    bool overflow = false;
+};
+/// Synchronous read of the counters (debug and the end-of-request summary).
+KvStreamCounters kv_stream_counters(const KvStreamMap& m);
+
+/// Bytes of one block (page) of K and V together.
+uint64_t kv_block_bytes(const QsaShapes& s, int fmt);
+
+}  // namespace strata::kernels

include/strata/kernels/native_ple_postops.hpp

@@ -27,4 +27,12 @@ void native_ple_postops(const float* projected_key, const float* hidden,
                         const PleWeights& weights, const NativePlePostopsBuffers& buffers,
                         void* stream);
 
+/// The prompt path: T tokens at once, element for element the arithmetic of T calls of native_ple_postops with
+/// the history advanced after each.  `key` (T x 10240) holds the projected keys and is normalized in place;
+/// `hidden` (T x 10240) becomes the result; `value` T x 2560; `query_norm` (T x 10240, the normalized hidden, then
+/// the normalized gated values), `gated` (T x 10240) and `gate` (T x 4) are scratch; `history` (9 rows per
+/// channel) is advanced past the T tokens.
+void native_ple_postops_batch(float* key, float* hidden, const float* value, float* history, const PleWeights& w,
+                              float* query_norm, float* gated, float* gate, int T, void* stream);
+
 } // namespace strata::kernels

include/strata/kernels/qsa.hpp

@@ -110,7 +110,9 @@ inline QsaShapes qsa_real_shapes() {
     QsaShapes s;
     s.n_head = 24; s.n_head_kv = 2; s.head_dim = 256; s.n_rot = 64;
     s.idx_n_head = 4; s.idx_dim = 128; s.idx_block = 4; s.idx_top_k = 2048;
-    s.page_size = 512;
+    // one page = one indexer block (4 cells): the granule KV streaming keeps resident (kv_stream.hpp). The readers
+    // resolve a row per cell anyway, so the page size costs them nothing (measured: identical output and speed).
+    s.page_size = 4;
     return s;
 }
 
@@ -353,8 +355,12 @@ void qsa_attend_step(const float* q, const uint16_t* k_scratch, const uint16_t*
                      const int32_t* step, int64_t max_ids, const QsaShapes& s, float* attn, float* weights,
                      void* stream);
 
-/// The cell's position comes from `step`, so this is the form a graph may contain.
+struct KvHostPools;   // kv_stream.hpp
+
+/// The cell's position comes from `step`, so this is the form a graph may contain. With a host copy (KV streaming)
+/// the cell is written there too, and to VRAM only if its block is resident.
 void kv_append_step(uint16_t* k_pool, uint16_t* v_pool, const int32_t* page_table, const int32_t* step,
-                    const float* kcur, const float* vcur, const QsaShapes& s, void* stream);
+                    const float* kcur, const float* vcur, const QsaShapes& s, void* stream,
+                    const KvHostPools* host = nullptr);
 
 }  // namespace strata::kernels

include/strata/kernels/qsa_decode_attn.hpp

@@ -27,6 +27,8 @@ struct QsaAttnPools {
     const int8_t* v_q = nullptr;
     const uint16_t* k_scale = nullptr;  ///< fp16 [page][kv_head][page_size][head_dim / 64]
     const uint16_t* v_scale = nullptr;
+    const uint8_t* k_q4 = nullptr;      ///< q4_0 block_q4_0 [page][kv_head][page_size][head_dim / 32 * 18]
+    const uint8_t* v_q4 = nullptr;
     const int32_t* page_table = nullptr;
 };
 

include/strata/kernels/sampler.hpp

@@ -1,21 +1,21 @@
 // include/strata/kernels/sampler.hpp - the sampler chain, host-callable (P2.S2).
 //
-//     penalties  ->  top_k  ->  top_p  ->  temperature  ->  pick
+//     penalties -> top_k -> top_p -> min_p -> temperature -> pick
 //
-// The ORDER is specified in docs/sampling.md, transcribed from llama.cpp's own chain.  Two facts there are
-// easy to get backwards: temperature comes AFTER the truncation filters, and penalties come AFTER temperature.
-//
-// PENALTIES ARE NOT IMPLEMENTED YET: they need the token history, which is not in this signature, so a caller
-// passing a non-zero `penalty_last_n` gets a loud failure rather than a silently unpenalised sample.
+// llama.cpp's default chain (issue #53): ONE penalties stage, first, and temperature AFTER the truncation filters;
+// top_p cuts before min_p.  `src/kernels/cuda/sampler.cu` has the details and `sampler_parity` pins them.
 #pragma once
 
+#include <algorithm>
+#include <cstddef>
 #include <cstdint>
 
 namespace strata::kernels {
 
 struct SamplerParams {
-    int top_k = 20;              // 0 disables the filter
+    int top_k = 20;              // sampled path: 1..64 as given; 0 (off) or more than 64 keep the widest list, 64
     float top_p = 0.95f;         // 1.0 disables the filter
+    float min_p = 0.0f;          // 0 disables; keeps tokens with p >= min_p * p_max (a prefix of the top_k list)
     float temperature = 1.0f;    // <= 0 means greedy
     int min_keep = 1;            // top_p keeps at least this many
     int penalty_last_n = 0;      // 0 disables; the window over `history` to count occurrences in
@@ -35,9 +35,29 @@ struct SamplerParams {
 // reported by whatever synchronising call happens NEXT, which will be somewhere else entirely and will name a
 // buffer that has nothing to do with it.  A sticky async fault does not know where it came from.
 //
-// `history` is (n_tokens, history_len) int32, the most recent tokens for each row with any unused slots set to
-// -1; only the last `p.penalty_last_n` of each row are counted.  Pass nullptr and 0 when no penalties apply.
+// `history` is (n_tokens, history_len) int32 - ONE ROW PER TOKEN ROW, at a stride of `history_len` - the most
+// recent tokens that row's pick follows, with any unused slots set to -1; only the last `p.penalty_last_n` of
+// each row are counted, and ids outside [0, n_vocab) are ignored.  Pass nullptr and 0 when no penalties apply.
+// A verify window's rows need DIFFERENT histories: row t follows the window's drafts 1..t (`penalty_rows`).
 void sample_tokens(const float* logits, int n_tokens, int n_vocab, const int* history, int history_len,
                    const SamplerParams& p, int* out, void* stream);
 
+// The penalty-history rows of a verify window, on the host: row t of `out` (T rows of `h` slots) is the last `h`
+// tokens of `tail[0..n_tail)` followed by `window[0..t]`, most recent LAST, -1 in the unused front slots.
+// `tail` is what the state consumed before the window, `window[0]` the fed-back token and `window[1..]` the
+// drafts: row t is exactly the history plain decode counts when it picks the token after `window[t]`, so drafting
+// cannot change which tokens are penalised.  Row 0 alone is the single row the engine staged before 0.1.19.
+inline void penalty_rows(const int32_t* tail, int64_t n_tail, const int32_t* window, int T, int h, int32_t* out) {
+    for (int t = 0; t < T; ++t) {
+        int32_t* row = out + (size_t) t * (size_t) h;
+        const int64_t avail = n_tail + t + 1;                      // tail + window[0..t]
+        const int take = (int) std::min<int64_t>(h, avail);
+        std::fill(row, row + (h - take), -1);
+        for (int j = 0; j < take; ++j) {
+            const int64_t i = avail - take + j;                     // index into (tail, window)
+            row[h - take + j] = i < n_tail ? tail[i] : window[i - n_tail];
+        }
+    }
+}
+
 }  // namespace strata::kernels

include/strata/prefill/gemm.hpp

@@ -36,6 +36,9 @@ public:
     void native(const uint16_t* X, int ggml_type, const void* W_blocks, float* Y, int64_t T, int64_t N, int64_t K,
                 int64_t ldy = 0, float beta = 0.0f);
 
+    /// Caller-owned buffers only: the scratch and workspace moved (the prompt path laid its buffers out again).
+    void rebind(uint16_t* scratch, int64_t scratch_elems, void* workspace, size_t ws_bytes);
+
     uint16_t* scratch() const { return scratch_; }
     int64_t scratch_elems() const { return scratch_elems_; }
     void* stream() const { return stream_; }

include/strata/prefill/kernels.hpp

@@ -5,6 +5,8 @@
 // the kernel walks the chunk's tokens in order inside one launch.
 #pragma once
 
+#include "strata/kernels/kv_stream.hpp"
+
 #include <cstdint>
 
 namespace strata::prefill {
@@ -35,7 +37,7 @@ void gdn_recurrence(float* state, const float* h, const float* gate, const float
 
 // ---- MoE
 /// softmax over 512, top-10 (ties to the lower id), weights renormalised over the ten (the native router).
-void route(const float* logits, int32_t* ids, float* weights, int64_t T, void* stream);
+void route(const float* logits, int32_t* ids, float* weights, int64_t T, int64_t n_expert, void* stream);
 /// Expert blob (Strata pack layout, Q2_0) -> BF16 matrices: gate/up interleaved [1280, 2560], down [2560, 640].
 void blob_dequant(const uint8_t* blob, uint16_t* gu16, uint16_t* down16, void* stream);
 /// h16[n, r] = fp16(silu(gu[n, 2r]) * gu[n, 2r + 1])   (the interleaved expert gate/up)
@@ -63,11 +65,14 @@ void gate_attn(const float* attn, const float* q_full, uint16_t* out16, int64_t
 /// K, V: [T, 2, 256].
 void kv_append(const float* K, const float* V, int64_t T, int64_t pos0, const int32_t* page_table, int64_t page_size,
                uint16_t* k_pool, uint16_t* v_pool, int8_t* k_q, int8_t* v_q, uint16_t* k_scale, uint16_t* v_scale,
-               void* stream);
+               void* stream, const strata::kernels::KvHostPools* host = nullptr,
+               const strata::kernels::KvHostPools* stage = nullptr);
 
 /// fp32 -> fp16 bits and fp32 -> bf16, n elements (the two activation images of the prompt GEMMs).
 void to_f16(const float* x, uint16_t* y, int64_t n, void* stream);
 void to_bf16(const float* x, uint16_t* y, int64_t n, void* stream);
+/// y = fp32(fp16(x)): what an FP16 store of x would read back (the FP16 indexer-key experiment)
+void round_f16(const float* x, float* y, int64_t n, void* stream);
 /// Expert blob -> FP16 (Q2_0 values are exact in FP16).
 void blob_dequant_f16(const uint8_t* blob, uint16_t* gu16, uint16_t* down16, void* stream);
 

include/strata/prefill/moe_mmq.hpp

@@ -0,0 +1,73 @@
+// include/strata/prefill/moe_mmq.hpp - prompt-speed plan step 2b: the prompt path's experts through llama.cpp's MMQ
+// kernels (ggml-cuda mmq.cuh, MIT): the weights stay quantized and the activations are rounded to q8_1, the
+// products run on int8 tensor cores.  The dequantize-to-FP16 + cuBLAS path wrote ~10 MB of FP16 per expert and
+// multiplied in FP16; this reads the ~1.4-2 MB expert once.  A group of experts is gathered into one buffer
+// (`gather_*`, one launch per expert as its blob arrives) and multiplied in one launch per product.
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+namespace strata::prefill::mmq {
+
+/// This build has the MMQ path (the ggml sources were available to the build).
+bool built();
+/// MMQ covers this ggml type (the i-quants and Q2_0 the packs use; IQ1_M is not covered).
+bool supported(int ggml_type);
+/// Bytes of one expert's gate+up ([2*n_ff, n_embd]) or down ([n_embd, n_ff]) weights in `ggml_type`.
+size_t matrix_bytes(int ggml_type, int64_t rows, int64_t cols);
+/// Bytes of `rows` activation rows of `cols` values quantized for MMQ (the row padded to 512 values).
+size_t q8_bytes(int64_t rows, int64_t cols);
+
+/// q8_1 activations for MMQ against weights of `ggml_type`: row i of the output is row ids[i] of x (or row i when
+/// ids is null); `x` has `ld` floats per row.
+void quantize(const float* x, const int32_t* ids, void* xq, int ggml_type, int64_t cols, int64_t ld, int64_t rows,
+              void* stream);
+
+/// One launch over n experts whose weights lie `expert_bytes` apart from `w`: for expert e, the activation rows
+/// [bounds[e], bounds[e+1]) of `xq` (bounds on the device, n+1 entries) times its [w_rows, w_cols] matrix into
+/// dst rows of the same indices (`ld_dst` floats apart, via `ids`: dst row = ids[row], an identity table works).
+/// `total_rows`: the rows of xq; `max_rows`: the most rows one expert has (the launch grid).
+struct Product {
+    const void* w = nullptr;
+    int type = -1;
+    int64_t w_rows = 0, w_cols = 0;
+    size_t expert_bytes = 0;
+    int n = 0;
+    const void* xq = nullptr;
+    const int32_t* bounds = nullptr;
+    const int32_t* ids = nullptr;
+    int64_t total_rows = 0, max_rows = 0;
+    float* dst = nullptr;
+    int64_t ld_dst = 0;
+};
+
+/// The launch context (llama.cpp's MMQ keeps a small scratch pool for its stream-k fixup).  One per prompt path.
+class Context {
+public:
+    Context();
+    ~Context();
+    Context(const Context&) = delete;
+    Context& operator=(const Context&) = delete;
+    void run(const Product& p, void* stream);
+
+private:
+    void* ctx_ = nullptr;
+};
+
+/// A GGUF-native expert (gate at `gate`, up at `up`, down at `down`, each its GGUF rows) into a group buffer's
+/// slot: gate rows then up rows at `gu_dst`, down at `d_dst`.
+void gather_native(const void* gate, const void* up, size_t gu_half_bytes, const void* down, size_t d_bytes,
+                   void* gu_dst, void* d_dst, void* stream);
+/// A Strata-pack Q2_0 expert blob (codes and fp16 scales in separate planes, gate/up rows interleaved) into GGUF
+/// Q2_0 blocks: gate/up [1280, 2560] at `gu_dst` (rows stay interleaved), down [2560, 640] at `d_dst`.  Same values.
+void gather_strata_q2(const uint8_t* blob, void* gu_dst, void* d_dst, void* stream);
+
+/// h[r, k] = silu(gate) * up of GU rows [2 n_ff wide]: interleaved (gate 2k, up 2k+1: the Strata pack) or split
+/// (gate k, up n_ff + k: GGUF).  FP32 out (the down product's quantizer reads floats).
+void swiglu(const float* gu, float* h, int64_t rows, int64_t n_ff, bool interleaved, void* stream);
+
+/// dst[i] = i for i < n (the identity row map MMQ's MoE mode writes through).
+void iota(int32_t* dst, int64_t n, void* stream);
+
+}  // namespace strata::prefill::mmq

include/strata/prefill/prefill.hpp

@@ -48,6 +48,16 @@ public:
               core::ExpertSource* src, const core::ExpertCache* cache, const int32_t* host_res, int64_t chunk,
               void* stream, std::string& err, void* borrow = nullptr, uint64_t borrow_bytes = 0);
 
+    /// With borrowed buffers: lay them out again for chunks of `chunk` tokens (at most `init`'s) in `borrow` - a
+    /// request lends only the slots its prompt needs.  The stream must be idle (between prompts).
+    bool relayout(int64_t chunk, void* borrow, uint64_t borrow_bytes, std::string& err);
+    int64_t chunk() const;
+
+    /// The share of the streamed experts' bytes DMA-able straight from pinned RAM (1 = all).  Sizes the streamed
+    /// ring (a big one only pays when the copy engine, not the host copies, is the limit); set before bytes_needed.
+    static void set_pinned_share(double share);
+    static double pinned_share();
+
     /// Device bytes `init` needs for a chunk of `chunk` tokens (what a borrowed region must hold).
     static uint64_t bytes_needed(const core::ModelGeometry& g, const core::SessionState& ss, int64_t chunk);
 
@@ -69,7 +79,20 @@ public:
     /// embedding where non-null (an image's <|image_pad|> cells).  Null (default): every position embeds its token.
     const float* const* embd_rows = nullptr;
 
+    /// LAYER SPLIT (multi-GPU): this prompt path runs layers [layer_begin, layer_end) (-1: to the last) on the
+    /// device `init` runs on.  A stage that does not start at layer 0 reads each chunk's residual rows from the
+    /// previous stage instead of embedding the tokens; a stage that does not end at the last layer copies its rows
+    /// to pinned host buffers (two, allocated by `init`) and runs `next` on them - on a thread, so the next stage
+    /// reads chunk c while this one reads chunk c + 1.  `on_chunk` belongs on the last stage.  Set before `init`.
+    void set_stage(int64_t layer_begin, int64_t layer_end, Prefill* next) {
+        stage_lb_ = layer_begin; stage_le_ = layer_end; next_ = next;
+    }
+
 private:
+    int64_t stage_lb_ = 0, stage_le_ = -1;
+    Prefill* next_ = nullptr;
+    const float* hand_in_ = nullptr;    ///< the previous stage's rows of the chunk being read (host, pinned)
+    bool carve(std::size_t T, void* alloc);   // the device buffers of a chunk (prefill.cpp's Alloc)
     struct Impl;
     std::unique_ptr<Impl> impl_;
     PrefillStats stats_;

include/strata/program/conv_cache.hpp

@@ -0,0 +1,43 @@
+// include/strata/program/conv_cache.hpp - the conversation cache's retention policy: which checkpoint
+// leaves when the cache is over its slot budget.
+//
+// The cache holds conversation checkpoints: snapshots of the running state (the GDN recurrence, the PLE
+// history, the QSA indexer tails - `ConvCheckpoint` in generate.cpp) that a request resumes its prompt
+// from instead of reading those tokens again.  The request path keeps only the checkpoints whose tokens
+// are a prefix of the current prompt - any other checkpoint's positional cells have been overwritten,
+// because the KV cache is one arena that holds one branch of history at a time - so at any moment the
+// retained checkpoints are a CHAIN: sorted by length, each a prefix of the next.  That is a radix cache's
+// tree collapsed onto the one branch of history the session can hold.
+//
+// The chain's root is therefore the deepest point every request so far has shared - in practice the end
+// of the system prompt, which every new chat of the same client mounts through - and it is exactly the
+// node a radix cache keeps alive while its leaves rotate.  First-in-first-out kept it only by accident of
+// being first: after `prompt_cache` newer checkpoints it was gone, and the next chat read the whole
+// prefix again at prefill speed.  So:
+//
+//   * the root is PINNED: dropping it frees one slot's ~118 MB and costs every future conversation that
+//     shares the prefix a full re-read of it (a 30K system prompt is ~30 s at ~1,000 tok/s);
+//   * every other slot rotates by least recent use.  A checkpoint's stamp advances when it is created,
+//     when a new checkpoint lands on its length, and when a request mounts through it.
+//
+// With a budget below two slots there is no room to keep root and leaf apart, so the pin switches off and
+// the oldest checkpoint leaves - the previous behaviour, one slot = the newest point only.
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+namespace strata::program::conv_cache {
+
+/// The index in `stamps` of the chain item to drop once the chain holds more than `cap` items.  `stamps`
+/// are the items' last-use stamps; the caller owns the chain and erases the returned index.  Pure and
+/// deterministic so conv_cache_test.cpp can walk the scenarios by hand.
+inline size_t eviction_victim(const uint64_t* stamps, size_t n, int64_t cap) {
+    if (cap < 2 || n < 2) return 0;   // no room for root and leaf: the pin is off, the oldest leaves
+    size_t v = 1;                     // the root (0) is pinned; least recent use among the rest
+    for (size_t i = 2; i < n; ++i)
+        if (stamps[i] < stamps[v]) v = i;
+    return v;
+}
+
+}  // namespace strata::program::conv_cache

include/strata/spec/draft_policy.hpp

@@ -0,0 +1,52 @@
+// include/strata/spec/draft_policy.hpp - per verify round: the MTP's window, or a lookup (suffix) window?
+//
+// The suffix drafter (prompt lookup) proposes the tokens that followed an earlier repeat of the context. Taken
+// whenever it proposes more than the MTP, it lost 2-8% on ordinary text: a long lookup window costs much more to
+// verify than the MTP's usual 3-4 tokens, and it is only worth that when enough of it is accepted. llama.cpp's
+// lookup decoding answers the same problem with confidence thresholds on its n-gram statistics; here the policy
+// learns both sides online and compares expected committed tokens per millisecond:
+//
+//   MTP     E = the mean tokens a window of that size has committed (EMA), at the measured cost of that size
+//   lookup  E(k) = 1 + q + q^2 + ... + q^k for k <= the proposal, q = the acceptance rate of lookup drafts whose
+//           match was about as long (4 buckets of match length, decayed counts), at the measured cost of k + 1
+//
+// and takes the lookup window only when its best E/cost beats the MTP's by `margin`. Costs are the measured round
+// times per window size (EMA; sizes not seen yet are scaled from seen ones by a prior shape), so the policy adapts
+// to the machine and the context length. It only chooses which drafts to verify: the output is unchanged.
+#pragma once
+
+#include <array>
+
+namespace strata::spec {
+
+class DraftPolicy {
+public:
+    static constexpr int kMaxT = 8;
+    static constexpr int kBuckets = 4;
+
+    explicit DraftPolicy(int max_t, double margin = 0.03);
+
+    struct Pick {
+        bool lookup = false;
+        int t = 1;                      // window size (1 + drafts)
+    };
+    /// `t_mtp`: the MTP's window; `lookup_k`: the lookup proposal's length (0 = none); `match`: its match length.
+    Pick choose(int t_mtp, int lookup_k, int match) const;
+    /// After the round: the window it used, the drafts accepted, and the round's time (verify + commit + draft).
+    void observe(bool lookup, int t, int accepted, int match, double round_ms);
+
+    double lookup_rate(int match) const;   // current q for a match length
+    double cost_ms(int t) const;           // measured or scaled round time of a window of t tokens
+
+private:
+    static int bucket(int match);
+    double mtp_tokens(int t) const;
+
+    int max_t_;
+    double margin_;
+    std::array<double, kMaxT + 1> cost_{}, cost_n_{};      // round ms by window size
+    std::array<double, kMaxT + 1> mtp_tok_{}, mtp_n_{};    // tokens committed by MTP windows of that size
+    std::array<double, kBuckets> ok_{}, bad_{};            // lookup drafts accepted / windows cut short, decayed
+};
+
+}  // namespace strata::spec

serve/frontend.py

@@ -129,6 +129,14 @@ def images_of(messages: list[dict]) -> list[str]:
             for item in m["content"] if item.get("type") == "image"]
 
 
+def _late_system_to_user(messages: list[dict]) -> list[dict]:
+    """The chat template takes a system message only at the start ("System message must be at the beginning").
+    Clients also send them mid-conversation - Claude Code's hook context as {"role": "system"} after the first user
+    turn, some OpenAI clients a late "developer" message (issue #56) - so those become user messages, in place:
+    merging them into the first one would change the prompt's start and cost the conversation cache every turn."""
+    return [dict(m, role="user") if m.get("role") == "system" and i > 0 else m for i, m in enumerate(messages)]
+
+
 def openai_to_messages(req: dict) -> tuple[list[dict], list[dict] | None, dict]:
     """OpenAI Chat Completions -> (template messages, template tools, template kwargs)."""
     messages = []
@@ -161,7 +169,7 @@ def openai_to_messages(req: dict) -> tuple[list[dict], list[dict] | None, dict]:
             kwargs = {"enable_thinking": False}
         elif k == "reasoning_effort" and "enable_thinking" not in kwargs:
             kwargs.update(effort_kwargs(v))
-    return messages, tools, kwargs
+    return _late_system_to_user(messages), tools, kwargs
 
 
 def anthropic_to_messages(req: dict) -> tuple[list[dict], list[dict] | None, dict]:
@@ -210,7 +218,7 @@ def anthropic_to_messages(req: dict) -> tuple[list[dict], list[dict] | None, dic
         kwargs.update(effort_kwargs(effort))
     elif isinstance(thinking, dict) and thinking.get("budget_tokens"):
         kwargs.update(budget_effort(thinking["budget_tokens"]))
-    return messages, tools, kwargs
+    return _late_system_to_user(messages), tools, kwargs
 
 
 # ------------------------------------------------------------------------------------------------ output parser

serve/index.html

@@ -1,152 +0,0 @@
-<!doctype html>
-<html lang="en">
-<head>
-<meta charset="utf-8">
-<meta name="viewport" content="width=device-width, initial-scale=1">
-<title>Strata</title>
-<style>
-  :root { --bg:#f6f7f9; --card:#fff; --fg:#1b1f24; --mute:#667085; --line:#e3e6ea; --acc:#2f6fed; --me:#e9f0ff; --ok:#1a7f37; }
-  @media (prefers-color-scheme: dark) {
-    :root { --bg:#111418; --card:#1a1e24; --fg:#e7e9ec; --mute:#9aa3ae; --line:#2b313a; --acc:#6b9bff; --me:#1f2b44; --ok:#3fb950; }
-  }
-  * { box-sizing: border-box; }
-  body { margin:0; background:var(--bg); color:var(--fg); font:15px/1.5 system-ui, -apple-system, "Segoe UI", sans-serif; }
-  main { max-width:860px; margin:0 auto; padding:16px; display:flex; flex-direction:column; gap:12px; min-height:100vh; }
-  .card { background:var(--card); border:1px solid var(--line); border-radius:12px; padding:14px 16px; }
-  h1 { font-size:20px; margin:0; display:flex; align-items:center; gap:8px; }
-  .dot { width:10px; height:10px; border-radius:50%; background:var(--mute); display:inline-block; }
-  .dot.on { background:var(--ok); }
-  .mute { color:var(--mute); font-size:13px; }
-  code { background:var(--bg); border:1px solid var(--line); border-radius:6px; padding:1px 6px; font-size:13px; }
-  .grid { display:grid; grid-template-columns: max-content 1fr; gap:4px 12px; margin-top:8px; font-size:14px; }
-  #log { flex:1; display:flex; flex-direction:column; gap:10px; overflow-y:auto; min-height:240px; }
-  .msg { padding:10px 12px; border-radius:10px; white-space:pre-wrap; word-wrap:break-word; }
-  .msg.user { background:var(--me); align-self:flex-end; max-width:85%; }
-  .msg.bot { background:var(--card); border:1px solid var(--line); }
-  .msg img { max-width:220px; max-height:160px; display:block; border-radius:6px; margin-bottom:6px; }
-  details { color:var(--mute); font-size:13px; margin-bottom:6px; }
-  form { display:flex; gap:8px; align-items:flex-end; flex-wrap:wrap; }
-  textarea { flex:1; min-width:200px; min-height:52px; max-height:200px; resize:vertical; padding:10px; border-radius:10px;
-             border:1px solid var(--line); background:var(--card); color:var(--fg); font:inherit; }
-  button, label.btn { border:1px solid var(--line); background:var(--card); color:var(--fg); border-radius:10px; padding:10px 14px;
-                      font:inherit; cursor:pointer; }
-  button.primary { background:var(--acc); border-color:var(--acc); color:#fff; }
-  button:disabled { opacity:.5; cursor:default; }
-  .row { display:flex; gap:14px; align-items:center; flex-wrap:wrap; font-size:13px; color:var(--mute); }
-  #att { font-size:13px; color:var(--mute); }
-</style>
-</head>
-<body>
-<main>
-  <div class="card">
-    <h1><span id="dot" class="dot"></span> Strata <span id="state" class="mute">connecting ...</span></h1>
-    <div class="grid">
-      <span class="mute">Model</span><span id="model">-</span>
-      <span class="mute">Context</span><span id="ctx">-</span>
-      <span class="mute">Images</span><span id="img">-</span>
-      <span class="mute">OpenAI API</span><span><code id="oai"></code> (any API key<span id="keynote"></span>)</span>
-      <span class="mute">Anthropic API</span><span><code id="ant"></code></span>
-    </div>
-    <p class="mute" style="margin:10px 0 0">Point any app with an "OpenAI-compatible" or "Anthropic" provider at these
-      addresses. Close the Strata window to stop the model.</p>
-  </div>
-
-  <div id="log" class="card"></div>
-
-  <div class="card">
-    <form id="f">
-      <textarea id="q" placeholder="Ask something ... (Enter sends, Shift+Enter = new line)"></textarea>
-      <label class="btn" id="attbtn" hidden>Picture<input id="file" type="file" accept="image/png,image/jpeg,image/gif,image/bmp" hidden></label>
-      <button class="primary" id="send">Send</button>
-      <button type="button" id="stop" disabled>Stop</button>
-    </form>
-    <div class="row" style="margin-top:8px">
-      <label>Thinking <select id="think" style="padding:4px;border-radius:8px;border:1px solid var(--line);background:var(--card);color:var(--fg)">
-        <option value="none">off - answer at once</option>
-        <option value="low">low</option>
-        <option value="medium" selected>medium</option>
-        <option value="high">high - best on hard questions, slowest</option>
-      </select></label>
-      <span id="att"></span>
-      <button type="button" id="clear" style="padding:4px 10px">New chat</button>
-      <input id="key" type="password" placeholder="API key" hidden style="padding:6px;border-radius:8px;border:1px solid var(--line)">
-    </div>
-  </div>
-</main>
-<script>
-const $ = id => document.getElementById(id);
-const base = location.origin;
-$("oai").textContent = base + "/v1";
-$("ant").textContent = base + "/v1/messages";
-let history = [], image = null, ctrl = null;
-
-async function status() {
-  try {
-    const h = await (await fetch("/health")).json();
-    $("dot").classList.add("on"); $("state").textContent = "running";
-    $("model").textContent = h.model; $("ctx").textContent = h.max_context.toLocaleString() + " tokens";
-    $("img").textContent = h.images ? "on" : "off (run setup again and answer yes to images)";
-    $("attbtn").hidden = !h.images;
-    if (h.api_key) { $("key").hidden = false; $("keynote").textContent = " - this server needs its key"; }
-  } catch (e) { $("state").textContent = "not reachable - is the Strata window still open?"; }
-}
-status();
-
-function bubble(cls, text) {
-  const d = document.createElement("div"); d.className = "msg " + cls; d.textContent = text || "";
-  $("log").appendChild(d); $("log").scrollTop = $("log").scrollHeight; return d;
-}
-
-$("file").onchange = () => {
-  const f = $("file").files[0]; if (!f) return;
-  const r = new FileReader();
-  r.onload = () => { image = r.result; $("att").textContent = "picture: " + f.name; };
-  r.readAsDataURL(f);
-};
-$("clear").onclick = () => { history = []; $("log").innerHTML = ""; };
-$("stop").onclick = () => ctrl && ctrl.abort();
-$("q").addEventListener("keydown", e => { if (e.key === "Enter" && !e.shiftKey) { e.preventDefault(); $("f").requestSubmit(); } });
-
-$("f").onsubmit = async e => {
-  e.preventDefault();
-  const text = $("q").value.trim(); if (!text) return;
-  $("q").value = "";
-  const content = image ? [{type: "image_url", image_url: {url: image}}, {type: "text", text}] : text;
-  const u = bubble("user", "");
-  if (image) { const im = document.createElement("img"); im.src = image; u.appendChild(im); }
-  u.appendChild(document.createTextNode(text));
-  history.push({role: "user", content});
-  image = null; $("att").textContent = ""; $("file").value = "";
-  const b = bubble("bot", ""); const det = document.createElement("details"); const sum = document.createElement("summary");
-  sum.textContent = "thinking"; det.appendChild(sum); const th = document.createElement("div"); det.appendChild(th);
-  const ans = document.createElement("div"); b.appendChild(ans);
-  $("send").disabled = true; $("stop").disabled = false; ctrl = new AbortController();
-  let answer = "";
-  try {
-    const headers = {"Content-Type": "application/json"};
-    if ($("key").value) headers["Authorization"] = "Bearer " + $("key").value;
-    const body = {model: "strata", messages: history, stream: true, max_tokens: 8192};
-    body.reasoning_effort = $("think").value;
-    const r = await fetch("/v1/chat/completions", {method: "POST", headers, body: JSON.stringify(body), signal: ctrl.signal});
-    if (!r.ok) { ans.textContent = "error: " + ((await r.json()).error || {}).message; throw 0; }
-    const rd = r.body.getReader(), dec = new TextDecoder(); let buf = "";
-    for (;;) {
-      const {value, done} = await rd.read(); if (done) break;
-      buf += dec.decode(value, {stream: true});
-      let i;
-      while ((i = buf.indexOf("\n\n")) >= 0) {
-        const line = buf.slice(0, i).trim(); buf = buf.slice(i + 2);
-        if (!line.startsWith("data: ") || line === "data: [DONE]") continue;
-        const d = JSON.parse(line.slice(6)).choices[0].delta;
-        if (d.reasoning_content) { if (!det.parentNode) b.insertBefore(det, ans); th.textContent += d.reasoning_content; }
-        if (d.content) { answer += d.content; ans.textContent = answer; }
-        $("log").scrollTop = $("log").scrollHeight;
-      }
-    }
-  } catch (err) { if (err && err.name === "AbortError") ans.textContent = answer + " [stopped]"; }
-  history.push({role: "assistant", content: answer});
-  $("send").disabled = false; $("stop").disabled = true; ctrl = null;
-};
-</script>
-</body>
-</html>

serve/mcp.py

@@ -0,0 +1,669 @@
+"""serve/mcp.py - tools from MCP (Model Context Protocol) servers for the web app's chats.
+
+The user lists servers in the run config (`"mcp_servers"`, or Claude Desktop's `"mcpServers"` block pasted as is) or
+in a separate file (`--mcp-config`).  Each one is either a program Strata starts (`command`, `args`, `env`, `cwd`:
+the stdio transport, newline-delimited JSON-RPC 2.0 on its stdin/stdout) or an address (`url`, `headers`: the
+Streamable HTTP transport, JSON-RPC POSTed, answered as JSON or as an event stream).  Both are implemented here with
+the standard library only; no MCP SDK is needed.
+
+`McpHub` owns every configured server: it starts them in the background when the server starts, offers their tools to
+the model as OpenAI function tools named `<server>__<tool>` (two servers may both have a `search`), and routes a call
+back to the server that has it.  A server that does not start is left out with one log line; the chat works without
+it.  A server that stops later (a crash) is started again on its next call.
+
+Tool results reach the model as text: the MCP content blocks joined, capped at `max_result_chars` (a directory tree or
+a web page can be megabytes, far more than a context holds) with a note saying so.  A failed call - the tool's own
+error, a crash, a timeout - becomes a result that starts with "error:", so the model can react instead of the chat
+breaking.
+"""
+from __future__ import annotations
+
+import collections
+import json
+import os
+import re
+import shutil
+import signal
+import subprocess
+import sys
+import threading
+import time
+import urllib.error
+import urllib.request
+
+PROTOCOL = "2025-06-18"               # the MCP revision Strata asks for; the server's answer is used as given
+DEFAULTS = {"timeout_s": 60.0, "max_result_chars": 20000, "max_rounds": 8, "start_timeout_s": 120.0}
+
+
+class McpError(RuntimeError):
+    """A server answered with an error, stopped, or could not be reached."""
+
+
+class McpTimeout(McpError):
+    pass
+
+
+class McpCancelled(McpError):
+    """The chat was stopped while the call ran."""
+
+
+class _Slot:
+    """One request waiting for its response."""
+
+    def __init__(self):
+        self.done = threading.Event()
+        self.msg = None
+        self.error = None
+
+
+def _wait(slot: _Slot, timeout: float, cancel: threading.Event | None, what: str):
+    """Wait for a response, checking the chat's stop button and the deadline a few times a second."""
+    deadline = time.monotonic() + timeout
+    while not slot.done.wait(0.05):
+        if cancel is not None and cancel.is_set():
+            raise McpCancelled("stopped")
+        if time.monotonic() > deadline:
+            raise McpTimeout(f"{what} timed out after {timeout:g} s")
+    if slot.error:
+        raise slot.error
+    msg = slot.msg
+    if "error" in msg:
+        err = msg["error"] if isinstance(msg["error"], dict) else {"message": str(msg["error"])}
+        raise McpError(f"{err.get('message') or 'error'} (code {err.get('code')})")
+    return msg.get("result") or {}
+
+
+# ------------------------------------------------------------------------------------------------ transports
+class StdioTransport:
+    """A server Strata starts: JSON-RPC messages one per line on its stdin and stdout; its stderr is its log."""
+    kind = "stdio"
+
+    def __init__(self, name: str, cfg: dict):
+        self.name, self.cfg = name, cfg
+        self.proc = None
+        self.lock = threading.Lock()                    # the pending requests
+        self.write_lock = threading.Lock()              # stdin; separate, so a full pipe can't block the reader
+        self.pending: dict[int, _Slot] = {}
+        self.next_id = 0
+        self.stderr = collections.deque(maxlen=40)       # its last log lines, for the error message when it fails
+        self.protocol = PROTOCOL
+
+    def start(self):
+        env = dict(os.environ)
+        env.update({str(k): str(v) for k, v in (self.cfg.get("env") or {}).items()})
+        command = str(self.cfg["command"])
+        # On Windows `npx`, `uvx` and friends are .cmd files that CreateProcess finds only with their extension
+        exe = shutil.which(command, path=env.get("PATH")) or command
+        extra = {}
+        if os.name == "nt":
+            extra["creationflags"] = getattr(subprocess, "CREATE_NO_WINDOW", 0)
+        else:
+            extra["start_new_session"] = True           # its own process group, so close() ends its children too
+        try:
+            self.proc = subprocess.Popen([exe, *[str(a) for a in self.cfg.get("args") or []]], stdin=subprocess.PIPE,
+                                         stdout=subprocess.PIPE, stderr=subprocess.PIPE, cwd=self.cfg.get("cwd") or None,
+                                         env=env, **extra)
+        except OSError as e:
+            raise McpError(f"could not start {command!r}: {e}") from None
+        threading.Thread(target=self._read, daemon=True).start()
+        threading.Thread(target=self._read_stderr, daemon=True).start()
+
+    def alive(self) -> bool:
+        return self.proc is not None and self.proc.poll() is None and not getattr(self, "ended", False)
+
+    @staticmethod
+    def _lines(f):
+        """A pipe's lines until it closes (or close() closed it under us)."""
+        try:
+            yield from f
+        except (OSError, ValueError):
+            return
+
+    def _read_stderr(self):
+        for raw in self._lines(self.proc.stderr):
+            line = raw.decode("utf-8", "replace").rstrip()
+            if line:
+                self.stderr.append(line)
+                if os.environ.get("STRATA_DEBUG"):
+                    print(f"[mcp {self.name}] {line}", flush=True)
+
+    def _read(self):
+        for raw in self._lines(self.proc.stdout):
+            line = raw.decode("utf-8", "replace").strip()
+            if not line:
+                continue
+            try:
+                msg = json.loads(line)
+            except ValueError:                           # a server that prints to stdout; the spec forbids it
+                self.stderr.append(line)
+                continue
+            for m in msg if isinstance(msg, list) else [msg]:
+                if isinstance(m, dict):
+                    self._dispatch(m)
+        self.ended = True
+        code = self.proc.poll()
+        err = McpError(f"the server stopped{f' (exit code {code})' if code is not None else ''}{self._tail()}")
+        with self.lock:
+            slots, self.pending = list(self.pending.values()), {}
+        for s in slots:
+            s.error = err
+            s.done.set()
+
+    def _tail(self) -> str:
+        return f": {self.stderr[-1][:300]}" if self.stderr else ""
+
+    def _dispatch(self, m: dict):
+        if "method" in m:
+            if "id" in m:                                # a request from the server: we offer no client features
+                try:
+                    if m["method"] == "ping":
+                        self._send({"jsonrpc": "2.0", "id": m["id"], "result": {}})
+                    else:
+                        self._send({"jsonrpc": "2.0", "id": m["id"],
+                                    "error": {"code": -32601, "message": f"method not supported: {m['method']}"}})
+                except McpError:
+                    pass
+            return                                       # notifications (log messages, list changes): not used
+        with self.lock:
+            slot = self.pending.pop(m.get("id"), None)
+        if slot is not None:                             # an answer to a call that timed out is simply dropped
+            slot.msg = m
+            slot.done.set()
+
+    def _send(self, msg: dict):
+        data = json.dumps(msg, ensure_ascii=False).encode("utf-8") + b"\n"
+        try:
+            with self.write_lock:
+                self.proc.stdin.write(data)
+                self.proc.stdin.flush()
+        except (OSError, ValueError):
+            raise McpError(f"the server stopped{self._tail()}") from None
+
+    def request(self, method: str, params: dict, timeout: float, cancel=None):
+        if not self.alive():
+            raise McpError(f"the server is not running{self._tail()}")
+        slot = _Slot()
+        with self.lock:
+            self.next_id += 1
+            rid = self.next_id
+            self.pending[rid] = slot
+        try:
+            self._send({"jsonrpc": "2.0", "id": rid, "method": method, "params": params})
+            return _wait(slot, timeout, cancel, method)
+        except (McpTimeout, McpCancelled) as e:
+            with self.lock:
+                self.pending.pop(rid, None)
+            try:                                         # tell the server to stop working on it
+                self.notify("notifications/cancelled", {"requestId": rid, "reason": str(e)})
+            except McpError:
+                pass
+            raise
+
+    def notify(self, method: str, params: dict | None = None):
+        self._send({"jsonrpc": "2.0", "method": method, **({"params": params} if params is not None else {})})
+
+    def close(self):
+        """As the spec says: close its stdin, give it a moment to exit, then end it (and what it started)."""
+        p = self.proc
+        if p is None:
+            return
+        try:
+            p.stdin.close()
+        except OSError:
+            pass
+        try:
+            p.wait(timeout=2)
+        except subprocess.TimeoutExpired:
+            self._kill_tree(p)
+        for f in (p.stdout, p.stderr):
+            try:
+                f.close()
+            except OSError:
+                pass
+
+    @staticmethod
+    def _kill_tree(p):
+        try:
+            if os.name == "nt":                          # npx.cmd -> node: end the whole tree
+                subprocess.run(["taskkill", "/PID", str(p.pid), "/T", "/F"], stdout=subprocess.DEVNULL,
+                               stderr=subprocess.DEVNULL, timeout=10)
+            else:
+                os.killpg(p.pid, signal.SIGTERM)
+                try:
+                    p.wait(timeout=2)
+                except subprocess.TimeoutExpired:
+                    os.killpg(p.pid, signal.SIGKILL)
+        except (OSError, subprocess.SubprocessError):
+            pass
+        try:
+            p.kill()
+        except OSError:
+            pass
+
+
+class HttpTransport:
+    """A server at an address (Streamable HTTP): every message is a POST; the answer is JSON or an event stream that
+    carries it.  The session id the server hands out at `initialize` rides on every later request."""
+    kind = "http"
+
+    def __init__(self, name: str, cfg: dict):
+        self.name, self.url = name, str(cfg["url"])
+        self.headers = {str(k): str(v) for k, v in (cfg.get("headers") or {}).items()}
+        self.session = None
+        self.protocol = None                             # sent as MCP-Protocol-Version once negotiated
+        self.next_id = 0
+        self.lock = threading.Lock()
+        self.broken = None
+
+    def start(self):
+        pass
+
+    def alive(self) -> bool:
+        return self.broken is None
+
+    def _headers(self) -> dict:
+        h = {"Content-Type": "application/json", "Accept": "application/json, text/event-stream",
+             "User-Agent": "strata", **self.headers}
+        if self.session:
+            h["Mcp-Session-Id"] = self.session
+        if self.protocol:
+            h["MCP-Protocol-Version"] = self.protocol
+        return h
+
+    def _post(self, msg: dict, rid, timeout: float, slot: _Slot | None):
+        try:
+            req = urllib.request.Request(self.url, data=json.dumps(msg, ensure_ascii=False).encode("utf-8"),
+                                         headers=self._headers(), method="POST")
+            with urllib.request.urlopen(req, timeout=timeout) as r:
+                if r.headers.get("Mcp-Session-Id"):
+                    self.session = r.headers["Mcp-Session-Id"]
+                if slot is None:                         # a notification: 202, nothing to read
+                    return
+                ctype = r.headers.get("Content-Type", "")
+                if "text/event-stream" in ctype:
+                    found = self._from_events(r, rid)
+                else:
+                    body = r.read()
+                    found = None
+                    parsed = json.loads(body) if body.strip() else None
+                    for m in parsed if isinstance(parsed, list) else [parsed]:
+                        if isinstance(m, dict) and m.get("id") == rid and "method" not in m:
+                            found = m
+                if found is None:
+                    raise McpError("the server's answer did not contain a response")
+                slot.msg = found
+        except urllib.error.HTTPError as e:
+            detail = ""
+            try:
+                detail = e.read().decode("utf-8", "replace")[:300]
+            except OSError:
+                pass
+            if e.code == 404 and self.session:           # the session expired: a restart starts a new one
+                self.broken = McpError("the server ended the session")
+            err = McpError(f"HTTP {e.code} {e.reason}{': ' + detail if detail else ''}")
+            if slot is None:
+                raise err from None
+            slot.error = err
+        except (OSError, ValueError, McpError) as e:
+            err = e if isinstance(e, McpError) else McpError(f"could not reach {self.url}: {e}")
+            if slot is None:
+                raise err from None
+            slot.error = err
+        finally:
+            if slot is not None:
+                slot.done.set()
+
+    @staticmethod
+    def _from_events(r, rid):
+        """Read the event stream until the event that answers `rid` (it may carry notifications first)."""
+        data = []
+        for raw in r:
+            line = raw.decode("utf-8", "replace").rstrip("\r\n")
+            if line.startswith("data:"):
+                data.append(line[5:].lstrip(" "))
+            elif not line and data:
+                try:
+                    m = json.loads("\n".join(data))
+                except ValueError:
+                    m = None
+                data = []
+                for x in m if isinstance(m, list) else [m]:
+                    if isinstance(x, dict) and x.get("id") == rid and "method" not in x:
+                        return x
+        return None
+
+    def request(self, method: str, params: dict, timeout: float, cancel=None):
+        if self.broken:
+            raise self.broken
+        with self.lock:
+            self.next_id += 1
+            rid = self.next_id
+        slot = _Slot()
+        # the POST runs on its own thread so the stop button and the deadline are noticed while it waits
+        threading.Thread(target=self._post, args=({"jsonrpc": "2.0", "id": rid, "method": method, "params": params},
+                                                  rid, timeout + 5, slot), daemon=True).start()
+        try:
+            return _wait(slot, timeout, cancel, method)
+        except (McpTimeout, McpCancelled) as e:
+            threading.Thread(target=self._quiet_notify, args=("notifications/cancelled",
+                                                              {"requestId": rid, "reason": str(e)}), daemon=True).start()
+            raise
+
+    def _quiet_notify(self, method, params):
+        try:
+            self.notify(method, params)
+        except McpError:
+            pass
+
+    def notify(self, method: str, params: dict | None = None):
+        self._post({"jsonrpc": "2.0", "method": method, **({"params": params} if params is not None else {})},
+                   None, 10, None)
+
+    def close(self):
+        if not self.session:
+            return
+        try:                                             # end the session on the server (it may not allow that)
+            req = urllib.request.Request(self.url, headers=self._headers(), method="DELETE")
+            urllib.request.urlopen(req, timeout=3).close()
+        except (OSError, ValueError):
+            pass
+
+
+# ------------------------------------------------------------------------------------------------ one server
+class McpServer:
+    """One configured server: started once (and again after it stopped), with the tools it listed."""
+
+    def __init__(self, name: str, cfg: dict, settings: dict):
+        self.name, self.cfg, self.settings = name, cfg, settings
+        self.kind = "http" if cfg.get("url") else "stdio"
+        self.status = "idle"             # idle -> starting -> ready | failed; ready -> stopped when it ends
+        self.error = None
+        self.tools: list[dict] = []
+        self.info: dict = {}
+        self.transport = None
+        self.lock = threading.Lock()
+        self.last_start = 0.0
+
+    def start(self) -> bool:
+        with self.lock:
+            return self._start()
+
+    def _start(self) -> bool:
+        if self.transport is not None:
+            self.transport.close()
+        self.status, self.error, self.last_start = "starting", None, time.monotonic()
+        t = HttpTransport(self.name, self.cfg) if self.kind == "http" else StdioTransport(self.name, self.cfg)
+        self.transport = t
+        timeout = float(self.settings["start_timeout_s"])
+        try:
+            t.start()
+            res = t.request("initialize", {"protocolVersion": PROTOCOL, "capabilities": {},
+                                           "clientInfo": {"name": "strata", "title": "Strata", "version": "1"}},
+                            timeout)
+            t.protocol = str(res.get("protocolVersion") or PROTOCOL)
+            self.info = {k: v for k, v in (res.get("serverInfo") or {}).items() if k in ("name", "title", "version")}
+            self.info["protocol"] = t.protocol
+            t.notify("notifications/initialized")
+            tools, cursor = [], None
+            for _ in range(100):                         # pages; a server that never stops paging is cut off
+                try:
+                    page = t.request("tools/list", {"cursor": cursor} if cursor else {}, timeout)
+                except McpError as e:
+                    if "code -32601" in str(e) and not tools:   # a server without tools (prompts/resources only)
+                        break
+                    raise
+                tools += [x for x in page.get("tools") or [] if isinstance(x, dict) and x.get("name")]
+                cursor = page.get("nextCursor")
+                if not cursor:
+                    break
+            self.tools, self.status = tools, "ready"
+            names = ", ".join(x["name"] for x in tools[:8]) + (", ..." if len(tools) > 8 else "")
+            print(f"[strata] MCP server {self.name!r}: {len(tools)} tool{'s' * (len(tools) != 1)}"
+                  f"{' (' + names + ')' if tools else ''}", flush=True)
+            return True
+        except McpError as e:
+            self.status, self.error, self.tools = "failed", str(e), []
+            t.close()
+            print(f"[strata] MCP server {self.name!r} did not start: {e}. Its tools are left out; the chat works "
+                  "without them.", flush=True)
+            return False
+
+    def call(self, tool: str, arguments: dict, timeout: float, cancel=None) -> dict:
+        with self.lock:
+            if self.transport is None or not self.transport.alive():
+                # it stopped since its last call (a crash, an expired session): start it again; one that failed to
+                # start is tried again at most every 10 s, so a broken command does not run on every call
+                if self.status == "failed" and time.monotonic() - self.last_start < 10:
+                    raise McpError(f"the server is not running ({self.error or 'it stopped'})")
+                print(f"[strata] MCP server {self.name!r} had stopped; starting it again", flush=True)
+                if not self._start():
+                    raise McpError(f"the server could not be started again: {self.error}")
+            t = self.transport
+        try:
+            return t.request("tools/call", {"name": tool, "arguments": arguments}, timeout, cancel)
+        except McpError as e:
+            if not t.alive():
+                self.status, self.error = "stopped", str(e)
+            raise
+
+    def close(self):
+        with self.lock:
+            if self.transport is not None:
+                self.transport.close()
+            if self.status in ("ready", "starting"):
+                self.status = "stopped"
+
+
+# ------------------------------------------------------------------------------------------------ the hub
+def result_text(result: dict) -> str:
+    """An MCP tool result as the text the model reads: the text blocks as they are, other blocks named."""
+    parts = []
+    for block in result.get("content") or []:
+        if not isinstance(block, dict):
+            continue
+        kind = block.get("type")
+        if kind == "text":
+            parts.append(str(block.get("text", "")))
+        elif kind == "resource":
+            res = block.get("resource") or {}
+            parts.append(str(res["text"]) if "text" in res else f"[resource {res.get('uri', '')} "
+                                                                  f"({res.get('mimeType', 'binary')}), not shown]")
+        elif kind == "resource_link":
+            parts.append(f"[resource link: {block.get('name') or ''} {block.get('uri', '')}]".replace("  ", " "))
+        elif kind in ("image", "audio"):
+            size = len(block.get("data") or "") * 3 // 4
+            parts.append(f"[{kind} ({block.get('mimeType', '?')}, {size:,} bytes): not shown to the model]")
+        else:
+            parts.append(json.dumps(block, ensure_ascii=False))
+    if not parts and result.get("structuredContent") is not None:
+        parts.append(json.dumps(result["structuredContent"], ensure_ascii=False))
+    return "\n".join(parts)
+
+
+def _clean(name: str) -> str:
+    return re.sub(r"[^A-Za-z0-9_-]", "_", str(name)) or "x"
+
+
+class McpHub:
+    """Every configured server, and the merged, namespaced tool list the model sees."""
+
+    def __init__(self, servers: dict[str, dict], settings: dict | None = None):
+        self.settings = {**DEFAULTS, **(settings or {})}
+        self.servers = {name: McpServer(name, cfg, self.settings) for name, cfg in servers.items()}
+        self.threads: list[threading.Thread] = []
+        self._routes: dict[str, tuple[McpServer, str]] = {}
+
+    def start(self, wait: bool = False):
+        """Start every server on its own thread (npx may download a package first: that must not delay the chat)."""
+        self.threads = [threading.Thread(target=s.start, daemon=True, name=f"mcp-{s.name}")
+                        for s in self.servers.values()]
+        for t in self.threads:
+            t.start()
+        if wait:
+            self.wait(None)
+
+    def wait(self, timeout: float | None) -> None:
+        """Wait (up to `timeout` s in all) for servers that are still starting."""
+        end = None if timeout is None else time.monotonic() + timeout
+        for t in self.threads:
+            t.join(None if end is None else max(0.0, end - time.monotonic()))
+
+    def routes(self) -> dict[str, tuple[McpServer, str]]:
+        """tool name the model sees -> (server, the tool's own name).  A server that stopped after it was ready keeps
+        its tools here: a call starts it again."""
+        out = {}
+        for s in self.servers.values():
+            if s.status not in ("ready", "stopped"):
+                continue
+            for t in s.tools:
+                name = f"{_clean(s.name)}__{_clean(t['name'])}"[:64]
+                n = 2
+                while name in out:                       # two names that differ only in cleaned characters
+                    name = f"{name[:60]}_{n}"
+                    n += 1
+                out[name] = (s, t["name"])
+        self._routes = out
+        return out
+
+    def template_tools(self, exclude=()) -> list[dict]:
+        """The tools in the chat template's form ({name, description, parameters}); `exclude`: names the request
+        brought itself (those win)."""
+        out = []
+        for name, (s, tool) in self.routes().items():
+            if name in exclude:
+                continue
+            t = next(x for x in s.tools if x["name"] == tool)
+            schema = t.get("inputSchema") if isinstance(t.get("inputSchema"), dict) else {}
+            out.append({"name": name, "description": str(t.get("description") or t.get("title") or ""),
+                        "parameters": schema or {"type": "object", "properties": {}}})
+        return out
+
+    def openai_tools(self) -> list[dict]:
+        return [{"type": "function", "function": t} for t in self.template_tools()]
+
+    def call(self, name: str, arguments: dict, cancel: threading.Event | None = None) -> dict:
+        """Run one tool -> {"ok", "text" (what the model reads, capped), "chars" (its full length), "truncated",
+        "ms", "server", "tool"}.  Errors become text starting with "error:"; only McpCancelled is raised."""
+        t0 = time.monotonic()
+        server, tool = self._routes.get(name) or self.routes().get(name, (None, name))
+        out = {"server": server.name if server else None, "tool": tool}
+        if server is None:
+            text, ok = f"error: there is no tool named {name!r}", False
+        else:
+            try:
+                res = server.call(tool, arguments if isinstance(arguments, dict) else {},
+                                  float(self.settings["timeout_s"]), cancel)
+                text, ok = result_text(res), not res.get("isError")
+                if not ok:
+                    text = "error: " + (text or "the tool reported an error")
+            except McpCancelled:
+                raise
+            except McpError as e:
+                text, ok = f"error: {e}", False
+        cap = int(self.settings["max_result_chars"])
+        full = len(text)
+        if cap > 0 and full > cap:
+            text = text[:cap] + (f"\n\n[... truncated: the tool returned {full:,} characters; only the first "
+                                 f"{cap:,} are shown]")
+        out.update(ok=ok, text=text, chars=full, truncated=cap > 0 and full > cap,
+                   ms=int((time.monotonic() - t0) * 1000))
+        return out
+
+    def status(self) -> dict:
+        """GET /mcp: every server, its state and its tools (for the web app)."""
+        routes = self.routes()
+        servers = []
+        for s in self.servers.values():
+            names = {tool: n for n, (srv, tool) in routes.items() if srv is s}
+            servers.append({"name": s.name, "transport": s.kind, "status": s.status, "error": s.error,
+                            "info": s.info,
+                            "tools": [{"name": names.get(t["name"], t["name"]), "tool": t["name"],
+                                       "description": str(t.get("description") or "")[:300]} for t in s.tools]})
+        return {"servers": servers, "tools": len(routes),
+                "settings": {k: self.settings[k] for k in ("timeout_s", "max_result_chars", "max_rounds")}}
+
+    def close(self):
+        for s in self.servers.values():
+            try:
+                s.close()
+            except Exception:  # noqa: BLE001 - closing is best effort
+                pass
+
+
+# ------------------------------------------------------------------------------------------------ config
+def _check_server(name, cfg, where) -> dict:
+    if not isinstance(cfg, dict):
+        raise SystemExit(f"[strata] {where}: MCP server {name!r} must be an object with \"command\" or \"url\"")
+    if cfg.get("url"):
+        if cfg.get("type") == "sse":
+            raise SystemExit(f"[strata] {where}: MCP server {name!r} uses the old SSE transport (\"type\": \"sse\"); "
+                             "Strata speaks Streamable HTTP - most servers offer it at /mcp")
+        if not isinstance(cfg.get("headers") or {}, dict):
+            raise SystemExit(f"[strata] {where}: MCP server {name!r}: \"headers\" must be an object")
+        return cfg
+    if not isinstance(cfg.get("command"), str) or not cfg["command"].strip():
+        raise SystemExit(f"[strata] {where}: MCP server {name!r} needs \"command\" (a program to start) or \"url\"")
+    if not isinstance(cfg.get("args") or [], list):
+        raise SystemExit(f"[strata] {where}: MCP server {name!r}: \"args\" must be a list")
+    if not isinstance(cfg.get("env") or {}, dict):
+        raise SystemExit(f"[strata] {where}: MCP server {name!r}: \"env\" must be an object")
+    return cfg
+
+
+def servers_from(block, where: str) -> dict[str, dict]:
+    """{"name": {...}} with the entries checked; `"disabled": true` leaves one out."""
+    if block is None:
+        return {}
+    if not isinstance(block, dict):
+        raise SystemExit(f"[strata] {where}: the MCP servers must be an object {{\"name\": {{...}}}}")
+    return {str(n): _check_server(n, c, where) for n, c in block.items()
+            if not (isinstance(c, dict) and c.get("disabled") is True)}
+
+
+def settings_from(cfg: dict) -> dict:
+    """The run config's optional `"mcp"` block: timeout_s, max_result_chars, max_rounds, start_timeout_s."""
+    out = {}
+    for key, value in (cfg.get("mcp") or {}).items():
+        number = isinstance(value, (int, float)) and not isinstance(value, bool)
+        if key in ("timeout_s", "start_timeout_s"):
+            if not number or value <= 0:
+                raise SystemExit(f"[strata] config mcp.{key}={value!r}: expected a number of seconds > 0")
+            out[key] = float(value)
+        elif key in ("max_result_chars", "max_rounds"):
+            if not number or value != int(value) or value < 1:
+                raise SystemExit(f"[strata] config mcp.{key}={value!r}: expected a positive integer")
+            out[key] = int(value)
+        else:
+            print(f"[strata] config mcp.{key}={value!r}: unknown key, ignored", flush=True)
+    return out
+
+
+def hub_from_config(cfg: dict, mcp_config_path: str | None = None) -> McpHub | None:
+    """The run config's `"mcp_servers"` (or `"mcpServers"`) plus the servers in the --mcp-config file (Claude
+    Desktop's format: {"mcpServers": {...}}); a name in both takes the file's entry.  None when there are none."""
+    servers = {}
+    servers.update(servers_from(cfg.get("mcp_servers"), "config mcp_servers"))
+    servers.update(servers_from(cfg.get("mcpServers"), "config mcpServers"))
+    if mcp_config_path:
+        try:
+            with open(mcp_config_path, encoding="utf-8-sig") as f:
+                data = json.load(f)
+        except (OSError, ValueError) as e:
+            raise SystemExit(f"[strata] --mcp-config {mcp_config_path}: {e}") from None
+        block = data.get("mcpServers", data.get("mcp_servers")) if isinstance(data, dict) else None
+        if block is None:
+            raise SystemExit(f"[strata] --mcp-config {mcp_config_path}: expected {{\"mcpServers\": {{...}}}}")
+        servers.update(servers_from(block, f"--mcp-config {mcp_config_path}"))
+    if not servers:
+        return None
+    return McpHub(servers, settings_from(cfg))
+
+
+if __name__ == "__main__":                               # python -m serve.mcp config.json: list what a config offers
+    hub = hub_from_config(json.load(open(sys.argv[1], encoding="utf-8-sig")) if len(sys.argv) > 1 else {},
+                          sys.argv[2] if len(sys.argv) > 2 else None)
+    if hub is None:
+        sys.exit("no MCP servers configured")
+    hub.start(wait=True)
+    print(json.dumps(hub.status(), indent=1))
+    hub.close()

serve/mcp_fake_server.py

@@ -0,0 +1,172 @@
+"""serve/mcp_fake_server.py - a small MCP server for serve/test_mcp.py (no MCP SDK, no network beyond localhost).
+
+    python serve/mcp_fake_server.py                  stdio: JSON-RPC lines on stdin/stdout
+    python serve/mcp_fake_server.py --http 8765      Streamable HTTP on 127.0.0.1:8765/mcp
+
+Tools: echo (returns its text), fail (a tool error, isError), sleep (answers after `seconds`), big (`n` characters),
+die (the process exits mid-call), add (two integers).  tools/list pages two tools at a time (`--page N`), so the
+client's pagination is exercised.  `--crash-at-start` exits before answering anything.  With FAKE_MCP_LOG set, the
+server appends what it saw (calls, cancellations) to that file, one JSON line each.
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+TOOLS = [
+    {"name": "echo", "description": "Returns the text it is given.",
+     "inputSchema": {"type": "object", "properties": {"text": {"type": "string"}}, "required": ["text"]}},
+    {"name": "fail", "description": "Always fails.", "inputSchema": {"type": "object", "properties": {}}},
+    {"name": "sleep", "description": "Waits, then answers.",
+     "inputSchema": {"type": "object", "properties": {"seconds": {"type": "number"}}}},
+    {"name": "big", "description": "Returns n characters.",
+     "inputSchema": {"type": "object", "properties": {"n": {"type": "integer"}}}},
+    {"name": "die", "description": "Ends the server.", "inputSchema": {"type": "object", "properties": {}}},
+    {"name": "add", "description": "Adds two integers.",
+     "inputSchema": {"type": "object", "properties": {"a": {"type": "integer"}, "b": {"type": "integer"}}}},
+]
+PAGE = 2
+cancelled: set = set()
+
+
+def log(event: dict):
+    path = os.environ.get("FAKE_MCP_LOG")
+    if path:
+        with open(path, "a", encoding="utf-8") as f:
+            f.write(json.dumps(event) + "\n")
+
+
+def text(t, error=False):
+    return {"content": [{"type": "text", "text": t}], "isError": error}
+
+
+def handle(msg: dict):
+    """-> the response for a request, None for a notification."""
+    method, rid, params = msg.get("method"), msg.get("id"), msg.get("params") or {}
+    if "id" not in msg:
+        if method == "notifications/cancelled":
+            cancelled.add(params.get("requestId"))
+            log({"cancelled": params.get("requestId")})
+        return None
+    if method == "initialize":
+        return {"jsonrpc": "2.0", "id": rid, "result": {
+            "protocolVersion": params.get("protocolVersion", "2025-06-18"), "capabilities": {"tools": {}},
+            "serverInfo": {"name": "fake", "version": "0.1"}}}
+    if method == "tools/list":
+        start = int(params.get("cursor") or 0)
+        page = {"tools": TOOLS[start:start + PAGE]}
+        if start + PAGE < len(TOOLS):
+            page["nextCursor"] = str(start + PAGE)
+        return {"jsonrpc": "2.0", "id": rid, "result": page}
+    if method == "tools/call":
+        name, args = params.get("name"), params.get("arguments") or {}
+        log({"call": name, "arguments": args})
+        if name == "echo":
+            result = text(str(args.get("text", "")))
+        elif name == "fail":
+            result = text("it failed on purpose", error=True)
+        elif name == "sleep":
+            time.sleep(float(args.get("seconds", 1)))
+            result = text("slept")
+        elif name == "big":
+            result = text("y" * int(args.get("n", 100000)))
+        elif name == "die":
+            sys.stderr.write("dying on purpose\n")
+            sys.stderr.flush()
+            os._exit(3)
+        elif name == "add":
+            result = text(str(int(args.get("a", 0)) + int(args.get("b", 0))))
+        else:
+            return {"jsonrpc": "2.0", "id": rid, "error": {"code": -32602, "message": f"unknown tool {name}"}}
+        return {"jsonrpc": "2.0", "id": rid, "result": result}
+    return {"jsonrpc": "2.0", "id": rid, "error": {"code": -32601, "message": f"no method {method}"}}
+
+
+def stdio():
+    for f in (sys.stdin, sys.stdout, sys.stderr):      # MCP's stdio is UTF-8, whatever the Windows code page is
+        f.reconfigure(encoding="utf-8")
+    out = threading.Lock()
+
+    def answer(msg):
+        r = handle(msg)
+        if r is not None and msg.get("id") not in cancelled:
+            with out:
+                sys.stdout.write(json.dumps(r) + "\n")
+                sys.stdout.flush()
+
+    for line in sys.stdin:
+        if not line.strip():
+            continue
+        msg = json.loads(line)
+        if msg.get("method") == "tools/call":           # calls on threads: a slow one must not block cancellation
+            threading.Thread(target=answer, args=(msg,), daemon=True).start()
+        else:
+            answer(msg)
+
+
+def http_server(port: int = 0) -> ThreadingHTTPServer:
+    """Streamable HTTP at /mcp: JSON answers, except tools/call, which answers as an event stream (a progress
+    notification first); a session id from initialize is required afterwards."""
+    sessions = set()
+
+    class H(BaseHTTPRequestHandler):
+        protocol_version = "HTTP/1.0"
+
+        def log_message(self, *a):
+            pass
+
+        def do_POST(self):
+            msg = json.loads(self.rfile.read(int(self.headers.get("Content-Length", 0))))
+            if msg.get("method") != "initialize" and self.headers.get("Mcp-Session-Id") not in sessions:
+                self.send_response(400)
+                self.end_headers()
+                return
+            r = handle(msg)
+            if r is None:
+                self.send_response(202)
+                self.end_headers()
+                return
+            if msg.get("method") == "tools/call":
+                self.send_response(200)
+                self.send_header("Content-Type", "text/event-stream")
+                self.end_headers()
+                note = {"jsonrpc": "2.0", "method": "notifications/progress", "params": {"progress": 1}}
+                self.wfile.write(b"event: message\ndata: " + json.dumps(note).encode() + b"\n\n")
+                self.wfile.write(b"event: message\ndata: " + json.dumps(r).encode() + b"\n\n")
+                return
+            body = json.dumps(r).encode()
+            self.send_response(200)
+            self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(body)))
+            if msg.get("method") == "initialize":
+                sid = f"s-{len(sessions) + 1}"
+                sessions.add(sid)
+                self.send_header("Mcp-Session-Id", sid)
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_DELETE(self):
+            sessions.discard(self.headers.get("Mcp-Session-Id"))
+            self.send_response(200)
+            self.end_headers()
+
+    httpd = ThreadingHTTPServer(("127.0.0.1", port), H)
+    threading.Thread(target=httpd.serve_forever, daemon=True).start()
+    return httpd
+
+
+if __name__ == "__main__":
+    if "--page" in sys.argv:
+        PAGE = int(sys.argv[sys.argv.index("--page") + 1])
+    if "--crash-at-start" in sys.argv:
+        sys.stderr.write("cannot start: missing configuration\n")
+        sys.exit(2)
+    if "--http" in sys.argv:
+        server = http_server(int(sys.argv[sys.argv.index("--http") + 1]))
+        print(f"listening on http://127.0.0.1:{server.server_address[1]}/mcp", flush=True)
+        threading.Event().wait()
+    stdio()

serve/telemetry.py

@@ -0,0 +1,247 @@
+"""serve/telemetry.py - hardware readings for the web app's Monitor tab (idea from PR #22 by code-martin).
+
+A background thread samples once a second and keeps the last 60 readings of each series for the sparklines:
+- GPU: NVIDIA's own NVML library (nvml.dll / libnvidia-ml.so.1, installed with every driver) through ctypes, so no
+  pip package is needed: load, VRAM, temperature, power, PCIe link and throughput.
+- CPU, RAM, disk: `psutil` when it is installed (setup installs it); without it the CPU and RAM readings fall back to
+  the OS (Windows GlobalMemoryStatusEx / GetSystemTimes, Linux /proc) and the disk rate is absent.
+Anything that cannot be read is None; nothing here can stop the server.
+"""
+from __future__ import annotations
+
+import collections
+import ctypes
+import os
+import platform
+import sys
+import threading
+import time
+
+HISTORY = 60
+
+
+# ------------------------------------------------------------------------------------------------ NVML
+class _Nvml:
+    class Util(ctypes.Structure):
+        _fields_ = [("gpu", ctypes.c_uint), ("memory", ctypes.c_uint)]
+
+    class Mem(ctypes.Structure):
+        _fields_ = [("total", ctypes.c_ulonglong), ("free", ctypes.c_ulonglong), ("used", ctypes.c_ulonglong)]
+
+    def __init__(self, index=0):
+        self.lib = self.dev = None
+        names = ["nvml.dll", os.path.join(os.environ.get("ProgramFiles", r"C:\Program Files"),
+                                          "NVIDIA Corporation", "NVSMI", "nvml.dll")] if os.name == "nt" \
+            else ["libnvidia-ml.so.1", "libnvidia-ml.so"]
+        for n in names:
+            try:
+                self.lib = ctypes.CDLL(n)
+                break
+            except OSError:
+                continue
+        if self.lib is None:
+            return
+        try:
+            init = getattr(self.lib, "nvmlInit_v2", None) or self.lib.nvmlInit
+            if init() != 0:
+                self.lib = None
+                return
+            h = ctypes.c_void_p()
+            get = getattr(self.lib, "nvmlDeviceGetHandleByIndex_v2", None) or self.lib.nvmlDeviceGetHandleByIndex
+            if get(ctypes.c_uint(index), ctypes.byref(h)) != 0:
+                self.lib = None
+                return
+            self.dev = h
+        except (AttributeError, OSError):
+            self.lib = None
+
+    def ok(self):
+        return self.lib is not None and self.dev is not None
+
+    def _uint(self, fn, *args):
+        v = ctypes.c_uint()
+        try:
+            return v.value if getattr(self.lib, fn)(self.dev, *args, ctypes.byref(v)) == 0 else None
+        except (AttributeError, OSError):
+            return None
+
+    def name(self):
+        buf = ctypes.create_string_buffer(96)
+        try:
+            if self.lib.nvmlDeviceGetName(self.dev, buf, ctypes.c_uint(96)) == 0:
+                return buf.value.decode(errors="replace")
+        except (AttributeError, OSError):
+            pass
+        return None
+
+    def read(self):
+        out = {}
+        u = self.Util()
+        try:
+            if self.lib.nvmlDeviceGetUtilizationRates(self.dev, ctypes.byref(u)) == 0:
+                out["util"] = u.gpu
+        except (AttributeError, OSError):
+            pass
+        m = self.Mem()
+        try:
+            if self.lib.nvmlDeviceGetMemoryInfo(self.dev, ctypes.byref(m)) == 0:
+                out["mem_used"], out["mem_total"] = m.used, m.total
+        except (AttributeError, OSError):
+            pass
+        out["temp"] = self._uint("nvmlDeviceGetTemperature", ctypes.c_uint(0))          # NVML_TEMPERATURE_GPU
+        mw = self._uint("nvmlDeviceGetPowerUsage")
+        out["power"] = mw / 1000.0 if mw is not None else None
+        lim = self._uint("nvmlDeviceGetEnforcedPowerLimit")
+        out["power_limit"] = lim / 1000.0 if lim is not None else None
+        out["pcie_gen"] = self._uint("nvmlDeviceGetCurrPcieLinkGeneration")        # drops at idle (power saving)
+        out["pcie_gen_max"] = self._uint("nvmlDeviceGetMaxPcieLinkGeneration")
+        out["pcie_width"] = self._uint("nvmlDeviceGetCurrPcieLinkWidth")
+        rx = self._uint("nvmlDeviceGetPcieThroughput", ctypes.c_uint(1))                 # NVML_PCIE_UTIL_RX_BYTES, KB/s
+        tx = self._uint("nvmlDeviceGetPcieThroughput", ctypes.c_uint(0))
+        out["pcie_rx_mb"] = rx / 1024.0 if rx is not None else None
+        out["pcie_tx_mb"] = tx / 1024.0 if tx is not None else None
+        return out
+
+
+# ------------------------------------------------------------------------------------------------ CPU / RAM
+def _cpu_name():
+    if os.name == "nt":
+        try:
+            import winreg
+            k = winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE, r"HARDWARE\DESCRIPTION\System\CentralProcessor\0")
+            return winreg.QueryValueEx(k, "ProcessorNameString")[0].strip()
+        except OSError:
+            pass
+    elif os.path.exists("/proc/cpuinfo"):
+        for line in open("/proc/cpuinfo", encoding="utf-8", errors="replace"):
+            if line.startswith("model name"):
+                return line.split(":", 1)[1].strip()
+    return platform.processor() or None
+
+
+class _CpuRamFallback:
+    """CPU load and RAM without psutil."""
+
+    def __init__(self):
+        self.prev = self._times()
+
+    def _times(self):
+        if os.name == "nt":
+            idle, kern, user = (ctypes.c_ulonglong() for _ in range(3))
+            if ctypes.windll.kernel32.GetSystemTimes(ctypes.byref(idle), ctypes.byref(kern), ctypes.byref(user)):
+                return idle.value, kern.value + user.value           # kernel time includes idle
+            return None
+        try:
+            f = [int(x) for x in open("/proc/stat").readline().split()[1:]]
+            return f[3] + f[4], sum(f)
+        except (OSError, ValueError):
+            return None
+
+    def cpu(self):
+        cur = self._times()
+        prev, self.prev = self.prev, cur
+        if not cur or not prev or cur[1] == prev[1]:
+            return None
+        return max(0.0, min(100.0, 100.0 * (1 - (cur[0] - prev[0]) / (cur[1] - prev[1]))))
+
+    @staticmethod
+    def ram():
+        if os.name == "nt":
+            class MS(ctypes.Structure):
+                _fields_ = [("dwLength", ctypes.c_ulong), ("dwMemoryLoad", ctypes.c_ulong),
+                            ("ullTotalPhys", ctypes.c_ulonglong), ("ullAvailPhys", ctypes.c_ulonglong),
+                            ("ullTotalPageFile", ctypes.c_ulonglong), ("ullAvailPageFile", ctypes.c_ulonglong),
+                            ("ullTotalVirtual", ctypes.c_ulonglong), ("ullAvailVirtual", ctypes.c_ulonglong),
+                            ("ullAvailExtendedVirtual", ctypes.c_ulonglong)]
+            m = MS()
+            m.dwLength = ctypes.sizeof(MS)
+            if ctypes.windll.kernel32.GlobalMemoryStatusEx(ctypes.byref(m)):
+                return m.ullTotalPhys - m.ullAvailPhys, m.ullTotalPhys
+            return None, None
+        try:
+            info = dict(line.split(":", 1) for line in open("/proc/meminfo"))
+            total = int(info["MemTotal"].split()[0]) * 1024
+            avail = int(info["MemAvailable"].split()[0]) * 1024
+            return total - avail, total
+        except (OSError, KeyError, ValueError):
+            return None, None
+
+
+# ------------------------------------------------------------------------------------------------ the sampler
+class Telemetry:
+    def __init__(self, extra=None, gpu_index=0):
+        """`extra()` -> dict of more series to record each second (the server's tok/s).  `gpu_index`: the card the
+        engine runs on, numbered as nvidia-smi and NVML number them (by PCI bus)."""
+        self.extra = extra
+        self.lock = threading.Lock()
+        self.now: dict = {}
+        self.hist = collections.defaultdict(lambda: collections.deque(maxlen=HISTORY))
+        self.gpu = _Nvml(gpu_index)
+        try:
+            import psutil  # noqa: F401
+            self.ps = sys.modules["psutil"]
+        except ImportError:
+            self.ps = None
+        self.fallback = _CpuRamFallback()
+        self.static = {
+            "gpu_name": self.gpu.name() if self.gpu.ok() else None,
+            "cpu_name": _cpu_name(),
+            "cores": (self.ps.cpu_count(logical=False) if self.ps else None) or None,
+            "threads": os.cpu_count(),
+            "psutil": self.ps is not None,
+        }
+        self._disk_prev = None
+        threading.Thread(target=self._loop, daemon=True).start()
+
+    def _disk(self):
+        if not self.ps:
+            return None, None
+        try:
+            c = self.ps.disk_io_counters()
+        except (OSError, RuntimeError):
+            return None, None
+        t = time.time()
+        prev, self._disk_prev = self._disk_prev, (t, c.read_bytes, c.write_bytes)
+        if prev is None or t <= prev[0]:
+            return None, None
+        dt = t - prev[0]
+        return (c.read_bytes - prev[1]) / dt / 2**20, (c.write_bytes - prev[2]) / dt / 2**20
+
+    def sample(self):
+        s = {}
+        if self.gpu.ok():
+            g = self.gpu.read()
+            s.update({f"gpu_{k}": v for k, v in g.items()})
+        if self.ps:
+            try:
+                s["cpu"] = self.ps.cpu_percent(interval=None)
+                vm = self.ps.virtual_memory()
+                s["ram_used"], s["ram_total"] = vm.total - vm.available, vm.total
+            except (OSError, RuntimeError):
+                pass
+        else:
+            s["cpu"] = self.fallback.cpu()
+            s["ram_used"], s["ram_total"] = self.fallback.ram()
+        s["disk_read_mb"], s["disk_write_mb"] = self._disk()
+        if self.extra:
+            try:
+                s.update(self.extra())
+            except Exception:  # noqa: BLE001 - telemetry must never take the server down
+                pass
+        return s
+
+    def _loop(self):
+        while True:
+            s = self.sample()
+            with self.lock:
+                self.now = s
+                for k in ("gpu_util", "gpu_mem_used", "gpu_temp", "gpu_power", "gpu_pcie_rx_mb", "cpu", "ram_used",
+                          "disk_read_mb", "tok_s"):
+                    v = s.get(k)
+                    self.hist[k].append(round(v, 2) if isinstance(v, float) else v)
+            time.sleep(1.0)
+
+    def snapshot(self):
+        with self.lock:
+            return {"now": dict(self.now), "history": {k: list(v) for k, v in self.hist.items()},
+                    "static": dict(self.static)}

serve/test_mcp.py

@@ -0,0 +1,434 @@
+"""serve/test_mcp.py - tools from MCP servers (serve/mcp.py) and the web app's tool loop, against the mock engine and
+the fake MCP server in serve/mcp_fake_server.py (no GPU, no pack, no MCP SDK).
+
+    python -m unittest serve.test_mcp -v
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+import tempfile
+import threading
+import time
+import unittest
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from serve import mcp_fake_server as fake  # noqa: E402
+from serve.frontend import ChatTemplate  # noqa: E402
+from serve.mcp import McpCancelled, McpHub, hub_from_config  # noqa: E402
+from serve.server import ByteTokenizer, MockEngine, Service, serve  # noqa: E402
+
+ROOT = Path(__file__).resolve().parents[1]
+FAKE = str(ROOT / "serve" / "mcp_fake_server.py")
+CTX = 8192
+
+
+def stdio(*args, **extra):
+    return {"command": sys.executable, "args": [FAKE, *args], **extra}
+
+
+def call_script(name, **params):
+    """The model's text for one tool call, as Qwen's template asks for it."""
+    body = "".join(f"<parameter={k}>\n{v}\n</parameter>\n" for k, v in params.items())
+    return f"</think>\n\nLet me check.\n\n<tool_call>\n<function={name}>\n{body}</function>\n</tool_call>"
+
+
+class ScriptedEngine(MockEngine):
+    """A mock engine with one script per call (the last one repeats): a tool call first, then the answer."""
+
+    def __init__(self, tok, scripts, max_context=CTX, delay_s=0.0):
+        super().__init__(tok, list(scripts), max_context=max_context, delay_s=delay_s)
+        self.prompts = []
+
+    def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+        self.prompts.append(list(ids))
+        yield from super().generate(ids, max_new, sampling, cancel, embeddings)
+
+    def prompt_text(self, i):
+        return bytes(t for t in self.prompts[i] if t < 256).decode("utf-8", "replace")
+
+
+# ------------------------------------------------------------------------------------------------ the client
+class Client(unittest.TestCase):
+    """The stdio client against the fake server: start, pagination, calls, errors, crashes, timeouts, the cap."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.log = tempfile.NamedTemporaryFile(delete=False, suffix=".jsonl")
+        cls.log.close()
+        os.environ["FAKE_MCP_LOG"] = cls.log.name
+        cls.hub = McpHub({"fake": stdio(), "paged": stdio("--page", "4"), "broken": stdio("--crash-at-start"),
+                          "missing": {"command": "strata-no-such-program"}},
+                         {"timeout_s": 1.5, "max_result_chars": 1000})
+        cls.hub.start(wait=True)
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.hub.close()
+        del os.environ["FAKE_MCP_LOG"]
+        os.unlink(cls.log.name)
+
+    def test_initialize_and_list_every_page(self):
+        s = self.hub.servers["fake"]
+        self.assertEqual(s.status, "ready")
+        self.assertEqual(s.info["name"], "fake")
+        self.assertEqual(s.info["protocol"], "2025-06-18")
+        want = [t["name"] for t in fake.TOOLS]
+        self.assertEqual([t["name"] for t in s.tools], want)         # 3 pages of 2
+        self.assertEqual([t["name"] for t in self.hub.servers["paged"].tools], want)   # 2 pages of 4
+
+    def test_namespaced_openai_tools(self):
+        tools = self.hub.openai_tools()
+        names = [t["function"]["name"] for t in tools]
+        self.assertIn("fake__echo", names)
+        self.assertIn("paged__echo", names)
+        echo = next(t for t in tools if t["function"]["name"] == "fake__echo")
+        self.assertEqual(echo["type"], "function")
+        self.assertEqual(echo["function"]["parameters"]["properties"]["text"]["type"], "string")
+
+    def test_a_server_that_fails_is_left_out(self):
+        for name, why in (("broken", "missing configuration"), ("missing", "could not start")):
+            with self.subTest(name=name):
+                s = self.hub.servers[name]
+                self.assertEqual(s.status, "failed")
+                self.assertIn(why, s.error)
+                self.assertFalse([t for t in self.hub.template_tools() if t["name"].startswith(name + "__")])
+        st = {s["name"]: s for s in self.hub.status()["servers"]}
+        self.assertEqual(st["broken"]["status"], "failed")
+        self.assertEqual(st["fake"]["status"], "ready")
+
+    def test_call(self):
+        self.hub.routes()
+        r = self.hub.call("fake__echo", {"text": "hello wörld"})
+        self.assertEqual((r["ok"], r["text"], r["server"], r["tool"]), (True, "hello wörld", "fake", "echo"))
+        self.assertEqual(self.hub.call("fake__add", {"a": 2, "b": 40})["text"], "42")
+
+    def test_errors_become_text(self):
+        self.hub.routes()
+        r = self.hub.call("fake__fail", {})                         # the tool's own error (isError)
+        self.assertFalse(r["ok"])
+        self.assertEqual(r["text"], "error: it failed on purpose")
+        r = self.hub.call("nobody__nothing", {})                    # not a tool at all
+        self.assertTrue(r["text"].startswith("error: there is no tool"))
+        # a JSON-RPC error from the server
+        s = self.hub.servers["fake"]
+        with self.assertRaises(Exception) as ctx:
+            s.call("no-such-tool", {}, 5)
+        self.assertIn("unknown tool", str(ctx.exception))
+
+    def test_timeout(self):
+        self.hub.routes()
+        t0 = time.monotonic()
+        r = self.hub.call("fake__sleep", {"seconds": 5})
+        self.assertLess(time.monotonic() - t0, 4)
+        self.assertFalse(r["ok"])
+        self.assertIn("timed out after 1.5 s", r["text"])
+        self.assertEqual(self.hub.call("fake__echo", {"text": "still here"})["text"], "still here")
+
+    def test_truncation(self):
+        self.hub.routes()
+        r = self.hub.call("fake__big", {"n": 50000})
+        self.assertTrue(r["ok"])
+        self.assertTrue(r["truncated"])
+        self.assertEqual(r["chars"], 50000)
+        self.assertTrue(r["text"].startswith("y" * 1000 + "\n\n[... truncated"))
+        self.assertIn("50,000 characters", r["text"])
+        self.assertLess(len(r["text"]), 1200)
+
+    def test_crash_mid_call_then_restart(self):
+        hub = McpHub({"solo": stdio()}, {"timeout_s": 5})
+        hub.start(wait=True)
+        try:
+            hub.routes()
+            r = hub.call("solo__die", {})
+            self.assertFalse(r["ok"])
+            self.assertIn("the server stopped", r["text"])
+            self.assertIn("dying on purpose", r["text"])
+            self.assertEqual(hub.servers["solo"].status, "stopped")
+            self.assertIn("solo__echo", hub.routes())                # its tools stay: the next call restarts it
+            self.assertEqual(hub.call("solo__echo", {"text": "back"})["text"], "back")
+            self.assertEqual(hub.servers["solo"].status, "ready")
+        finally:
+            hub.close()
+
+    def test_cancel(self):
+        self.hub.routes()
+        cancel = threading.Event()
+        threading.Timer(0.3, cancel.set).start()
+        t0 = time.monotonic()
+        with self.assertRaises(McpCancelled):
+            self.hub.call("fake__sleep", {"seconds": 1.2}, cancel)
+        self.assertLess(time.monotonic() - t0, 1.0)
+        time.sleep(0.3)
+        seen = [json.loads(line) for line in Path(self.log.name).read_text().splitlines()]
+        self.assertTrue(any("cancelled" in e for e in seen), seen)   # notifications/cancelled reached the server
+
+
+class HttpClient(unittest.TestCase):
+    """Streamable HTTP: JSON answers, an event-stream answer, the session id."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.httpd = fake.http_server(0)
+        url = f"http://127.0.0.1:{cls.httpd.server_address[1]}/mcp"
+        cls.hub = McpHub({"web": {"url": url}, "down": {"url": "http://127.0.0.1:9/mcp"}}, {"timeout_s": 1.5})
+        cls.hub.start(wait=True)
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.hub.close()
+        cls.httpd.shutdown()
+        cls.httpd.server_close()
+
+    def test_list_and_call(self):
+        s = self.hub.servers["web"]
+        self.assertEqual(s.status, "ready", s.error)
+        self.assertEqual(len(s.tools), len(fake.TOOLS))             # pages, with the session id on each request
+        self.assertEqual(s.transport.session, "s-1")
+        self.hub.routes()
+        self.assertEqual(self.hub.call("web__echo", {"text": "over http"})["text"], "over http")   # event stream
+        self.assertEqual(self.hub.call("web__fail", {})["text"], "error: it failed on purpose")
+
+    def test_timeout_and_unreachable(self):
+        self.hub.routes()
+        r = self.hub.call("web__sleep", {"seconds": 4})
+        self.assertIn("timed out", r["text"])
+        self.assertEqual(self.hub.servers["down"].status, "failed")
+        self.assertIn("could not reach", self.hub.servers["down"].error)
+
+
+class Config(unittest.TestCase):
+    def test_both_spellings_and_the_file(self):
+        with tempfile.TemporaryDirectory() as d:
+            path = os.path.join(d, "claude_desktop_config.json")
+            Path(path).write_text(json.dumps({"mcpServers": {"b": {"url": "http://x/mcp"}, "a": stdio()}}),
+                                  encoding="utf-8")
+            hub = hub_from_config({"mcp_servers": {"a": {"command": "old"}},
+                                   "mcpServers": {"c": stdio(), "off": {"command": "x", "disabled": True}},
+                                   "mcp": {"timeout_s": 5, "max_result_chars": 100, "max_rounds": 3}}, path)
+            self.assertEqual(sorted(hub.servers), ["a", "b", "c"])
+            self.assertEqual(hub.servers["a"].cfg["command"], sys.executable)   # the file wins a name clash
+            self.assertEqual(hub.servers["b"].kind, "http")
+            self.assertEqual((hub.settings["timeout_s"], hub.settings["max_result_chars"], hub.settings["max_rounds"]),
+                             (5.0, 100, 3))
+        self.assertIsNone(hub_from_config({}))
+        self.assertEqual(McpHub({}).settings["max_result_chars"], 20000)
+
+    def test_bad_entries_stop_the_start(self):
+        for cfg in ({"mcp_servers": {"x": {}}}, {"mcp_servers": {"x": {"command": "a", "args": "b"}}},
+                    {"mcp_servers": ["x"]}, {"mcpServers": {"x": {"url": "http://a", "type": "sse"}}},
+                    {"mcp_servers": {"x": {"command": "a"}}, "mcp": {"timeout_s": 0}},
+                    {"mcp_servers": {"x": {"command": "a"}}, "mcp": {"max_rounds": 1.5}}):
+            with self.subTest(cfg=cfg), self.assertRaises(SystemExit):
+                hub_from_config(cfg)
+        with self.assertRaises(SystemExit):
+            hub_from_config({}, os.path.join(tempfile.gettempdir(), "strata-no-such-mcp.json"))
+
+
+# ------------------------------------------------------------------------------------------------ the tool loop
+class ToolLoop(unittest.TestCase):
+    """The web app's chat (`"strata_mcp": true`): the model calls an MCP tool, the server runs it and the model
+    answers with its result; plain API requests never see the MCP tools."""
+
+    def setUp(self):
+        self.log = tempfile.NamedTemporaryFile(delete=False, suffix=".jsonl")
+        self.log.close()
+        os.environ["FAKE_MCP_LOG"] = self.log.name
+        self.hub = McpHub({"fake": stdio(), "broken": stdio("--crash-at-start")},
+                          {"timeout_s": 10, "max_result_chars": 500, "max_rounds": 2})
+        self.hub.start(wait=True)
+        self.httpd = None
+
+    def tearDown(self):
+        if self.httpd:
+            self.httpd.shutdown()
+            self.httpd.server_close()
+        self.hub.close()
+        del os.environ["FAKE_MCP_LOG"]
+        os.unlink(self.log.name)
+
+    def start(self, *scripts, delay_s=0.0):
+        tok = ByteTokenizer()
+        self.engine = ScriptedEngine(tok, list(scripts), delay_s=delay_s)
+        self.svc = Service(self.engine, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        self.svc.mcp = self.hub
+        self.httpd = serve(self.svc, port=0)
+        self.base = f"http://127.0.0.1:{self.httpd.server_address[1]}"
+
+    def post(self, body, headers=None, stream=True):
+        body = {"model": "m", "messages": [{"role": "user", "content": "check it"}], "stream": stream, **body}
+        req = urllib.request.Request(self.base + "/v1/chat/completions", data=json.dumps(body).encode(),
+                                     headers={"Content-Type": "application/json", **(headers or {})})
+        try:
+            with urllib.request.urlopen(req, timeout=30) as r:
+                text = r.read().decode()
+                return r.status, text
+        except urllib.error.HTTPError as e:
+            with e:
+                return e.code, e.read().decode()
+
+    @staticmethod
+    def chunks(text):
+        return [json.loads(line[6:]) for line in text.splitlines() if line.startswith("data: {")]
+
+    def test_tool_call_then_answer(self):
+        self.start(call_script("fake__echo", text="hello from the tool"), "</think>\n\nThe tool said hello.")
+        code, text = self.post({"strata_mcp": True})
+        self.assertEqual(code, 200, text)
+        cs = self.chunks(text)
+        mcp = [c["strata_mcp"] for c in cs if "strata_mcp" in c]
+        self.assertEqual([m["event"] for m in mcp], ["start", "call", "result"])
+        self.assertEqual(mcp[1]["arguments"], {"text": "hello from the tool"})
+        self.assertEqual((mcp[1]["server"], mcp[1]["tool"]), ("fake", "echo"))
+        self.assertEqual((mcp[2]["ok"], mcp[2]["text"]), (True, "hello from the tool"))
+        self.assertEqual(len({m["id"] for m in mcp}), 1)
+        content = "".join((c["choices"][0]["delta"].get("content") or "") for c in cs)
+        self.assertEqual(content, "Let me check.The tool said hello.")
+        self.assertFalse([c for c in cs if c["choices"][0]["delta"].get("tool_calls")])   # nothing for the client to run
+        self.assertEqual(cs[-1]["choices"][0]["finish_reason"], "stop")
+        self.assertEqual(len(self.engine.prompts), 2)
+        first, second = self.engine.prompt_text(0), self.engine.prompt_text(1)
+        self.assertIn('"name": "fake__echo"', first)                 # the MCP tools are in the prompt
+        self.assertNotIn("broken__", first)                          # the server that failed is left out
+        self.assertIn("<function=fake__echo>\n<parameter=text>\nhello from the tool\n</parameter>", second)
+        self.assertIn("<tool_response>\nhello from the tool\n</tool_response>", second)
+        self.assertEqual(cs[-1]["usage"]["prompt_tokens"], len(self.engine.prompts[1]))
+
+    def test_non_stream(self):
+        self.start(call_script("fake__add", a=2, b=3), "</think>\n\n5.")
+        code, text = self.post({"strata_mcp": True}, stream=False)
+        self.assertEqual(code, 200, text)
+        msg = json.loads(text)["choices"][0]["message"]
+        self.assertEqual(msg["content"], "Let me check.5.")
+        self.assertEqual(msg["strata_mcp"][-1]["text"], "5")
+
+    def test_plain_requests_get_no_mcp_tools(self):
+        self.start(call_script("fake__echo", text="x"), "</think>\n\nnever")
+        code, text = self.post({})
+        self.assertEqual(code, 200, text)
+        self.assertNotIn("fake__echo", self.engine.prompt_text(0))
+        cs = self.chunks(text)
+        self.assertFalse([c for c in cs if "strata_mcp" in c])
+        self.assertEqual(cs[-1]["choices"][0]["finish_reason"], "tool_calls")   # returned to the client as always
+        self.assertEqual(len(self.engine.prompts), 1)
+        self.assertNotIn('"call"', Path(self.log.name).read_text())    # nothing ran
+        for req in ({"strata_mcp": "yes"}, {"strata_mcp": 1}):         # only a real true opts in
+            self.post(req)
+            self.assertNotIn("fake__echo", self.engine.prompt_text(-1))
+
+    def test_own_tools_still_go_to_the_client(self):
+        own = [{"type": "function", "function": {"name": "lookup", "parameters": {"type": "object", "properties": {
+            "q": {"type": "string"}}}}}]
+        self.start(call_script("lookup", q="x"), "</think>\n\nnever")
+        code, text = self.post({"strata_mcp": True, "tools": own})
+        cs = self.chunks(text)
+        calls = [tc for c in cs for tc in c["choices"][0]["delta"].get("tool_calls") or []]
+        self.assertEqual(calls[0]["function"]["name"], "lookup")
+        self.assertEqual(cs[-1]["choices"][0]["finish_reason"], "tool_calls")
+        self.assertIn('"name": "lookup"', self.engine.prompt_text(0))
+        self.assertIn('"name": "fake__echo"', self.engine.prompt_text(0))
+        self.assertEqual(len(self.engine.prompts), 1)
+
+    def test_tool_error_is_a_result(self):
+        self.start(call_script("fake__fail"), "</think>\n\nIt failed.")
+        code, text = self.post({"strata_mcp": True})
+        mcp = [c["strata_mcp"] for c in self.chunks(text) if "strata_mcp" in c]
+        self.assertEqual((mcp[-1]["ok"], mcp[-1]["text"]), (False, "error: it failed on purpose"))
+        self.assertIn("<tool_response>\nerror: it failed on purpose\n</tool_response>", self.engine.prompt_text(1))
+        self.assertEqual(self.chunks(text)[-1]["choices"][0]["finish_reason"], "stop")
+
+    def test_truncated_for_the_model(self):
+        self.start(call_script("fake__big", n=3000), "</think>\n\nLong.")
+        self.post({"strata_mcp": True})
+        second = self.engine.prompt_text(1)
+        self.assertIn("y" * 500 + "\n\n[... truncated: the tool returned 3,000 characters", second)
+        self.assertNotIn("y" * 501, second)
+
+    def test_max_rounds(self):
+        self.start(call_script("fake__echo", text="again"))          # the model never stops calling
+        code, text = self.post({"strata_mcp": True})
+        mcp = [c["strata_mcp"] for c in self.chunks(text) if "strata_mcp" in c]
+        self.assertEqual(len(self.engine.prompts), 3)                # 2 rounds of tools, then the limit
+        self.assertEqual(sum(m["event"] == "call" for m in mcp), 2)
+        self.assertIn({"event": "limit", "max_rounds": 2}, mcp)
+        self.assertTrue(mcp[-1].get("skipped"))
+        self.assertEqual(self.chunks(text)[-1]["choices"][0]["finish_reason"], "stop")
+
+    def test_stop_during_a_tool(self):
+        """Closing the connection while a slow tool runs stops the tool (notifications/cancelled) and the loop."""
+        self.start(call_script("fake__sleep", seconds=8), "</think>\n\nnever")
+        body = {"model": "m", "messages": [{"role": "user", "content": "x"}], "stream": True, "strata_mcp": True}
+        req = urllib.request.Request(self.base + "/v1/chat/completions", data=json.dumps(body).encode(),
+                                     headers={"Content-Type": "application/json"})
+        r = urllib.request.urlopen(req, timeout=30)
+        for raw in r:
+            if b'"event": "call"' in raw:
+                break
+        r.close()                                                    # the web app's Stop aborts the fetch
+        t0 = time.monotonic()
+        while time.monotonic() - t0 < 6 and "cancelled" not in Path(self.log.name).read_text():
+            time.sleep(0.1)
+        self.assertIn("cancelled", Path(self.log.name).read_text())
+        self.assertLess(time.monotonic() - t0, 6)
+        time.sleep(0.5)
+        self.assertEqual(len(self.engine.prompts), 1)                # no second round
+        self.assertFalse(self.svc.status["busy"])
+
+    def test_only_from_the_app_s_own_page(self):
+        self.start("</think>\n\nhi")
+        req = urllib.request.Request(self.base + "/v1/chat/completions", data=json.dumps(
+            {"model": "m", "messages": [{"role": "user", "content": "x"}], "strata_mcp": True}).encode(),
+            headers={"Content-Type": "text/plain"})                  # a cross-site "simple" request
+        with self.assertRaises(urllib.error.HTTPError) as ctx:
+            urllib.request.urlopen(req, timeout=10)
+        self.assertEqual(ctx.exception.code, 415)
+        code, _ = self.post({"strata_mcp": True}, {"Origin": "http://evil.example"})
+        self.assertEqual(code, 403)
+        self.assertEqual(self.engine.prompts, [])
+        host = self.base.split("://", 1)[1]
+        code, _ = self.post({"strata_mcp": True}, {"Origin": "http://" + host})
+        self.assertEqual(code, 200)
+
+    def test_get_mcp(self):
+        self.start("</think>\n\nhi")
+        with urllib.request.urlopen(self.base + "/mcp", timeout=10) as r:
+            st = json.loads(r.read())
+        servers = {s["name"]: s for s in st["servers"]}
+        self.assertEqual(servers["fake"]["status"], "ready")
+        self.assertEqual(servers["fake"]["transport"], "stdio")
+        self.assertIn("fake__echo", [t["name"] for t in servers["fake"]["tools"]])
+        self.assertEqual(servers["broken"]["status"], "failed")
+        self.assertIn("missing configuration", servers["broken"]["error"])
+        self.assertEqual(st["tools"], len(fake.TOOLS))
+        self.assertEqual(st["settings"]["max_rounds"], 2)
+        self.svc.api_key = "k"
+        with self.assertRaises(urllib.error.HTTPError):
+            urllib.request.urlopen(self.base + "/mcp", timeout=10)
+        self.svc.api_key = ""
+
+
+class NoServers(unittest.TestCase):
+    def test_opt_in_without_servers_is_a_plain_chat(self):
+        tok = ByteTokenizer()
+        svc = Service(ScriptedEngine(tok, ["</think>\n\nhi"]), tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        httpd = serve(svc, port=0)
+        try:
+            base = f"http://127.0.0.1:{httpd.server_address[1]}"
+            with urllib.request.urlopen(base + "/mcp", timeout=10) as r:
+                self.assertEqual(json.loads(r.read()), {"servers": [], "tools": 0})
+            req = urllib.request.Request(base + "/v1/chat/completions", data=json.dumps(
+                {"model": "m", "messages": [{"role": "user", "content": "x"}], "strata_mcp": True}).encode(),
+                headers={"Content-Type": "application/json"})
+            with urllib.request.urlopen(req, timeout=10) as r:
+                self.assertEqual(json.loads(r.read())["choices"][0]["message"]["content"], "hi")
+        finally:
+            httpd.shutdown()
+            httpd.server_close()
+
+
+if __name__ == "__main__":
+    unittest.main()

serve/test_server.py

@@ -0,0 +1,530 @@
+"""serve/test_server.py - the max tokens budget over both APIs, against the mock engine (no GPU, no pack).
+
+    python -m unittest serve.test_server -v
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+import unittest
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from serve.frontend import ChatTemplate  # noqa: E402
+from serve.server import CTX_SLACK, ByteTokenizer, EngineDied, MockEngine, Service, StrataEngine, serve  # noqa: E402
+
+ROOT = Path(__file__).resolve().parents[1]
+CTX = 4096
+ANSWER = "x" * 2000                              # longer than the old 1024 fallback: one token per byte
+
+
+class RecordingEngine(MockEngine):
+    def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+        self.last_max_new = max_new
+        yield from super().generate(ids, max_new, sampling, cancel, embeddings)
+
+
+class MaxTokens(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        tok = ByteTokenizer()
+        cls.engine = RecordingEngine(tok, "</think>\n\n" + ANSWER, max_context=CTX)
+        cls.svc = Service(cls.engine, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        cls.httpd = serve(cls.svc, port=0)
+        cls.base = f"http://127.0.0.1:{cls.httpd.server_address[1]}"
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.httpd.shutdown()
+        cls.httpd.server_close()
+
+    def post(self, path, body):
+        req = urllib.request.Request(self.base + path, data=json.dumps(body).encode(),
+                                     headers={"Content-Type": "application/json"})
+        try:
+            with urllib.request.urlopen(req, timeout=30) as r:
+                return r.status, json.loads(r.read())
+        except urllib.error.HTTPError as e:
+            with e:
+                return e.code, json.loads(e.read())
+
+    def call(self, api, text="hi", **budget):
+        """-> (status, body, prompt tokens, completion tokens); `budget` is merged into the request as given."""
+        msgs = [{"role": "user", "content": text}]
+        if api == "openai":
+            s, b = self.post("/v1/chat/completions", {"model": "m", "messages": msgs, **budget})
+            u = b.get("usage", {})
+            return s, b, u.get("prompt_tokens"), u.get("completion_tokens")
+        s, b = self.post("/v1/messages", {"model": "m", "messages": msgs, **budget})
+        u = b.get("usage", {})
+        return s, b, u.get("input_tokens"), u.get("output_tokens")
+
+    def test_unset_budget_is_the_rest_of_the_context(self):
+        cases = {"openai": [{"max_tokens": -1}, {"max_tokens": 0}, {}, {"max_tokens": None},
+                            {"max_completion_tokens": -1}, {"max_completion_tokens": None, "max_tokens": None}],
+                 "anthropic": [{"max_tokens": -1}, {"max_tokens": 0}, {}, {"max_tokens": None}]}
+        for api, budgets in cases.items():
+            for budget in budgets:
+                with self.subTest(api=api, budget=budget):
+                    s, b, pt, ct = self.call(api, **budget)
+                    self.assertEqual(s, 200, b)
+                    self.assertEqual(self.engine.last_max_new, CTX - CTX_SLACK - pt)
+                    self.assertGreater(ct, 1024)          # the whole answer, not cut at the old 1024 fallback
+
+    def test_explicit_budget_is_honoured(self):
+        for api, budget in [("openai", {"max_tokens": 50}), ("openai", {"max_completion_tokens": 50}),
+                            ("openai", {"max_completion_tokens": 50, "max_tokens": 9}),
+                            ("anthropic", {"max_tokens": 50}), ("openai", {"max_tokens": 1500}),
+                            ("anthropic", {"max_tokens": 1500})]:
+            with self.subTest(api=api, budget=budget):
+                want = budget.get("max_completion_tokens") or budget["max_tokens"]
+                s, b, _, ct = self.call(api, **budget)
+                self.assertEqual(s, 200, b)
+                self.assertEqual(self.engine.last_max_new, want)
+                self.assertEqual(ct, want)
+
+    def test_explicit_budget_over_the_context_is_rejected(self):
+        for api in ("openai", "anthropic"):
+            with self.subTest(api=api):
+                s, b, _, _ = self.call(api, max_tokens=CTX)
+                self.assertEqual(s, 400)
+                self.assertIn("exceeds the context", b["error"]["message"])
+
+    def test_unset_budget_with_a_near_full_prompt(self):
+        _, _, pt0, _ = self.call("openai", max_tokens=1)
+        overhead = pt0 - len("hi")                  # the template's tokens around the user text
+        for api in ("openai", "anthropic"):
+            _, _, pa, _ = self.call(api, max_tokens=1)
+            over = pa - pt0                          # the Anthropic template may differ slightly
+            with self.subTest(api=api, room=5):     # a few tokens left: the budget is exactly those
+                text = "y" * (CTX - CTX_SLACK - overhead - over - 5)
+                s, b, pt, ct = self.call(api, text=text, max_tokens=-1)
+                self.assertEqual(s, 200, b)
+                self.assertEqual(self.engine.last_max_new, 5)
+                self.assertEqual(ct, 5)
+            with self.subTest(api=api, room=0):     # nothing left: rejected, not truncated
+                text = "y" * (CTX - CTX_SLACK - overhead - over)
+                s, b, _, _ = self.call(api, text=text)
+                self.assertEqual(s, 400, b)
+                self.assertIn("no room to answer", b["error"]["message"])
+
+    def test_debug_log_shows_the_resolved_budget(self):
+        import contextlib
+        import io
+        os.environ["STRATA_DEBUG"] = "1"
+        try:
+            for api in ("openai", "anthropic"):
+                with self.subTest(api=api):
+                    out = io.StringIO()
+                    with contextlib.redirect_stdout(out):
+                        _, _, pt, _ = self.call(api, max_tokens=-1)
+                    self.assertIn(f"max_new={CTX - CTX_SLACK - pt} ", out.getvalue())
+        finally:
+            del os.environ["STRATA_DEBUG"]
+
+
+class FitMaxTokens(unittest.TestCase):
+    """PR #24: --fit-max-tokens clamps an explicit budget that overshoots the context instead of a 400."""
+
+    @classmethod
+    def setUpClass(cls):
+        tok = ByteTokenizer()
+        cls.engine = RecordingEngine(tok, "</think>\n\n" + ANSWER, max_context=CTX)
+        cls.svc = Service(cls.engine, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"), fit_max_tokens=True)
+        cls.httpd = serve(cls.svc, port=0)
+        cls.base = f"http://127.0.0.1:{cls.httpd.server_address[1]}"
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.httpd.shutdown()
+        cls.httpd.server_close()
+
+    post = MaxTokens.post
+    call = MaxTokens.call
+
+    def test_overshoot_is_clamped_to_the_room(self):
+        for api in ("openai", "anthropic"):
+            with self.subTest(api=api):
+                s, b, pt, ct = self.call(api, max_tokens=CTX)
+                self.assertEqual(s, 200, b)
+                self.assertEqual(self.engine.last_max_new, CTX - CTX_SLACK - pt)
+
+    def test_a_budget_that_fits_is_unchanged(self):
+        s, b, _, ct = self.call("openai", max_tokens=50)
+        self.assertEqual(s, 200, b)
+        self.assertEqual(self.engine.last_max_new, 50)
+
+    def test_no_room_is_still_a_400(self):
+        _, _, pt0, _ = self.call("openai", max_tokens=1)
+        overhead = pt0 - len("hi")
+        s, b, _, _ = self.call("openai", text="y" * (CTX - CTX_SLACK - overhead), max_tokens=100)
+        self.assertEqual(s, 400, b)
+        self.assertIn("no room to answer", b["error"]["message"])
+
+
+class ClientShapes(unittest.TestCase):
+    """What real clients send: Claude Code posts /v1/messages?beta=true (issue #55) and puts hook context into the
+    conversation as a mid-conversation system message (issue #56); some OpenAI clients send a late developer message."""
+
+    @classmethod
+    def setUpClass(cls):
+        tok = ByteTokenizer()
+        cls.engine = RecordingPrompt(tok, "</think>\n\n2", max_context=CTX)
+        cls.svc = Service(cls.engine, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        cls.httpd = serve(cls.svc, port=0)
+        cls.base = f"http://127.0.0.1:{cls.httpd.server_address[1]}"
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.httpd.shutdown()
+        cls.httpd.server_close()
+
+    def post(self, path, body):
+        req = urllib.request.Request(self.base + path, data=json.dumps(body).encode(),
+                                     headers={"Content-Type": "application/json", "anthropic-version": "2023-06-01"})
+        try:
+            with urllib.request.urlopen(req, timeout=30) as r:
+                return r.status, json.loads(r.read())
+        except urllib.error.HTTPError as e:
+            with e:
+                return e.code, json.loads(e.read())
+
+    def prompt_text(self):
+        return bytes(i for i in self.engine.last_ids if i < 256).decode("utf-8", "replace")
+
+    def test_query_string(self):
+        body = {"model": "x", "max_tokens": 20, "messages": [{"role": "user", "content": "hi"}]}
+        for path in ("/v1/messages?beta=true", "/v1/chat/completions?api-version=1", "/v1/messages/?beta=true"):
+            status, b = self.post(path, body)
+            self.assertEqual(status, 200, (path, b))
+        status, _ = self.post("/v1/nothing?beta=true", body)
+        self.assertEqual(status, 404)
+
+    def test_anthropic_mid_conversation_system(self):
+        status, b = self.post("/v1/messages?beta=true", {
+            "model": "x", "max_tokens": 50,
+            "system": [{"type": "text", "text": "You are terse."}],
+            "messages": [
+                {"role": "user", "content": [{"type": "text", "text": "1+1? digits only"}]},
+                {"role": "system", "content": [{"type": "text", "text": "<system-reminder>answer in digits</system-reminder>"}]}]})
+        self.assertEqual(status, 200, b)
+        text = self.prompt_text()
+        self.assertIn("You are terse.", text)
+        self.assertIn("<system-reminder>answer in digits</system-reminder>", text)
+        self.assertLess(text.index("You are terse."), text.index("1+1?"))          # the first system stays first
+        self.assertLess(text.index("1+1?"), text.index("answer in digits"))        # the late one stays in place
+
+    def test_openai_late_developer_and_system(self):
+        status, b = self.post("/v1/chat/completions", {
+            "model": "x", "max_tokens": 50,
+            "messages": [{"role": "system", "content": "Be brief."}, {"role": "user", "content": "hello"},
+                         {"role": "assistant", "content": "hi"}, {"role": "developer", "content": "Now use digits."},
+                         {"role": "system", "content": "Also this."}, {"role": "user", "content": "1+1?"}]})
+        self.assertEqual(status, 200, b)
+        text = self.prompt_text()
+        for part in ("Be brief.", "Now use digits.", "Also this.", "1+1?"):
+            self.assertIn(part, text)
+
+    def test_leading_system_unchanged(self):
+        from serve.frontend import anthropic_to_messages, openai_to_messages
+        msgs, _, _ = openai_to_messages({"messages": [{"role": "developer", "content": "D"}, {"role": "user", "content": "u"}]})
+        self.assertEqual([m["role"] for m in msgs], ["system", "user"])
+        msgs, _, _ = anthropic_to_messages({"system": "S", "messages": [{"role": "user", "content": "u"}]})
+        self.assertEqual([m["role"] for m in msgs], ["system", "user"])
+
+
+class SamplingKeys(unittest.TestCase):
+    """The GEN line's sampling keys: top_k 0 ("off") or wider than the engine's 64 get the widest list, 64 (they used
+    to fall back to the engine default 20); a penalty always carries its window."""
+
+    def keys(self, **sampling):
+        return StrataEngine.sampling_keys(sampling).split()
+
+    def test_top_k(self):
+        self.assertIn("top_k=10", self.keys(temperature=0.7, top_k=10))
+        self.assertIn("top_k=64", self.keys(temperature=0.7, top_k=64))
+        self.assertIn("top_k=64", self.keys(temperature=0.7, top_k=0))
+        self.assertIn("top_k=64", self.keys(temperature=0.7, top_k=100))
+        for bad in (-1, True, 2.5, "20"):
+            self.assertFalse([k for k in self.keys(temperature=0.7, top_k=bad) if k.startswith("top_k=")], bad)
+
+    def test_tune_keys(self):
+        k = self.keys(temperature=0, strata_tune={"pcie_frac": 0.2, "spec_min_p": 0.7})
+        self.assertIn("pcie_frac=0.2", k)
+        self.assertIn("spec_min_p=0.7", k)
+        bad = self.keys(strata_tune={"pcie_frac": 3, "spec_min_p": True, "pool_workers": 2})
+        self.assertFalse([x for x in bad if x.split("=")[0] in ("pcie_frac", "spec_min_p", "pool_workers")])
+
+    def test_penalty_window(self):
+        self.assertIn("penalty_last_n=64", self.keys(presence_penalty=1.5))
+        self.assertIn("penalty_last_n=4096", self.keys(repetition_penalty=1.1, penalty_last_n=4096))
+        self.assertFalse([k for k in self.keys(temperature=0.7) if k.startswith("penalty")])
+
+
+class GpuChoice(unittest.TestCase):
+    """Issue #51: the config's \"gpu\" reaches the engine as CUDA_VISIBLE_DEVICES, numbered like nvidia-smi."""
+
+    def test_env(self):
+        from serve.server import child_env
+        env = child_env({"gpu": 1})
+        self.assertEqual(env["CUDA_VISIBLE_DEVICES"], "1")
+        self.assertEqual(env["CUDA_DEVICE_ORDER"], "PCI_BUS_ID")
+        plain = child_env({})                     # no choice: the environment as it was (existing installs)
+        self.assertEqual(plain.get("CUDA_VISIBLE_DEVICES"), os.environ.get("CUDA_VISIBLE_DEVICES"))
+        self.assertEqual(plain.get("CUDA_DEVICE_ORDER"), os.environ.get("CUDA_DEVICE_ORDER"))
+
+
+class RecordingPrompt(MockEngine):
+    def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+        self.last_ids = list(ids)
+        yield from super().generate(ids, max_new, sampling, cancel, embeddings)
+
+
+class DyingEngine(MockEngine):
+    """Issue #27: an engine that dies after a few tokens of its first answer, and comes back when restarted."""
+
+    def __init__(self, tok, script, max_context):
+        super().__init__(tok, script, max_context=max_context)
+        self.dead, self.restarts, self.die_after = False, 0, 5
+
+    def alive(self):
+        return not self.dead
+
+    def restart(self):
+        self.dead, self.die_after = False, None
+        self.restarts += 1
+
+    def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+        for i, t in enumerate(super().generate(ids, max_new, sampling, cancel, embeddings)):
+            if self.die_after is not None and i == self.die_after:
+                self.dead = True
+                raise EngineDied("the engine stopped unexpectedly (exit code -9)")
+            yield t
+
+
+class EngineDeath(unittest.TestCase):
+    """Issue #27: a dead engine is an error (not "length"), and the next request starts it again."""
+
+    def test_error_then_restart(self):
+        tok = ByteTokenizer()
+        eng = DyingEngine(tok, "</think>\n\n" + ANSWER, max_context=CTX)
+        svc = Service(eng, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        httpd = serve(svc, port=0)
+        base = f"http://127.0.0.1:{httpd.server_address[1]}"
+        try:
+            def post(body):
+                req = urllib.request.Request(base + "/v1/chat/completions", data=json.dumps(body).encode(),
+                                             headers={"Content-Type": "application/json"})
+                try:
+                    with urllib.request.urlopen(req, timeout=30) as r:
+                        return r.status, r.read().decode()
+                except urllib.error.HTTPError as e:
+                    with e:
+                        return e.code, e.read().decode()
+            msgs = [{"role": "user", "content": "hi"}]
+            code, text = post({"model": "m", "messages": msgs, "max_tokens": 50, "stream": True})
+            self.assertEqual(code, 200)
+            self.assertIn('"error"', text)
+            self.assertIn("stopped unexpectedly", text)
+            self.assertTrue(text.rstrip().endswith("data: [DONE]"))
+            self.assertEqual(svc.metrics()["requests"][0]["finish"], "error")
+            code, text = post({"model": "m", "messages": msgs, "max_tokens": 50})
+            self.assertEqual(code, 200, text)
+            self.assertEqual(eng.restarts, 1)
+            self.assertEqual(json.loads(text)["usage"]["completion_tokens"], 50)
+        finally:
+            httpd.shutdown()
+            httpd.server_close()
+
+    def test_engine_err_mid_stream(self):
+        """The engine's ERR line after the stream started reaches the client as an error event (it used to be a
+        400 written into the open stream, which clients read as an empty answer)."""
+        class ErrEngine(MockEngine):
+            def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+                yield None                                  # a prompt-progress heartbeat: the stream has started
+                raise ValueError("verify: layer 31 never rang (an illegal memory access was encountered)")
+
+        tok = ByteTokenizer()
+        svc = Service(ErrEngine(tok, ANSWER, max_context=CTX), tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        httpd = serve(svc, port=0)
+        base = f"http://127.0.0.1:{httpd.server_address[1]}"
+        try:
+            for path, body in [("/v1/chat/completions", {"model": "m", "stream": True, "max_tokens": 20,
+                                                          "messages": [{"role": "user", "content": "hi"}]}),
+                               ("/v1/messages", {"model": "m", "stream": True, "max_tokens": 20,
+                                                 "messages": [{"role": "user", "content": "hi"}]})]:
+                req = urllib.request.Request(base + path, data=json.dumps(body).encode(),
+                                             headers={"Content-Type": "application/json"})
+                with urllib.request.urlopen(req, timeout=30) as r:
+                    text = r.read().decode()
+                self.assertIn("illegal memory access", text, path)
+                self.assertNotIn("HTTP/1", text, path)
+                self.assertEqual(svc.metrics()["requests"][0]["finish"], "error")
+        finally:
+            httpd.shutdown()
+            httpd.server_close()
+
+
+class SharedSettings(unittest.TestCase):
+    """The web app's "Use for other apps too": POST /settings makes its Chat settings every client's defaults."""
+
+    @classmethod
+    def setUpClass(cls):
+        import tempfile
+
+        class Sampled(RecordingEngine):
+            def generate(self, ids, max_new, sampling, cancel, embeddings=None):
+                self.last_sampling = dict(sampling or {})
+                yield from super().generate(ids, max_new, sampling, cancel, embeddings)
+
+        tok = ByteTokenizer()
+        cls.engine = Sampled(tok, "</think>\n\nhello", max_context=CTX)
+        cls.svc = Service(cls.engine, tok, ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        cls.tmp = tempfile.TemporaryDirectory()
+        cls.svc.shared_path = os.path.join(cls.tmp.name, "strata-x.shared-settings.json")
+        cls.httpd = serve(cls.svc, port=0)
+        cls.base = f"http://127.0.0.1:{cls.httpd.server_address[1]}"
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.httpd.shutdown()
+        cls.httpd.server_close()
+        cls.tmp.cleanup()
+
+    def req(self, path, body, headers=None, raw=None):
+        h = {"Content-Type": "application/json", **(headers or {})}
+        r = urllib.request.Request(self.base + path, data=raw if raw is not None else json.dumps(body).encode(), headers=h)
+        try:
+            with urllib.request.urlopen(r, timeout=30) as resp:
+                return resp.status, json.loads(resp.read())
+        except urllib.error.HTTPError as e:
+            with e:
+                return e.code, json.loads(e.read())
+
+    def chat(self, **extra):
+        return self.req("/v1/chat/completions", {"model": "m", "messages": [{"role": "user", "content": "hi"}], **extra})
+
+    def tearDown(self):
+        self.svc.set_shared(None)
+
+    def test_other_apps_get_the_chat_settings(self):
+        d = {"temperature": 0.3, "top_p": 0.9, "top_k": 10, "seed": 7, "max_tokens": 77,
+             "reasoning_effort": "low", "experimental_speed_projection": False}
+        code, b = self.req("/settings", {"defaults": d})
+        self.assertEqual(code, 200, b)
+        self.assertTrue(b["shared"])
+        self.assertTrue(os.path.exists(self.svc.shared_path))
+        code, _ = self.chat()                                        # a client that sets nothing
+        self.assertEqual(code, 200)
+        got = self.engine.last_sampling
+        for k in ("temperature", "top_p", "top_k", "seed", "experimental_speed_projection"):
+            self.assertEqual(got[k], d[k], k)
+        self.assertEqual(self.engine.last_max_new, 77)
+        code, _ = self.chat(temperature=0.9, max_tokens=5)          # its own values win
+        self.assertEqual(self.engine.last_sampling["temperature"], 0.9)
+        self.assertEqual(self.engine.last_max_new, 5)
+        r = self.svc.with_shared({"messages": []}, "openai")
+        self.assertEqual(r["reasoning_effort"], "low")
+        self.assertEqual(self.svc.with_shared({"reasoning_effort": "high"}, "openai")["reasoning_effort"], "high")
+        self.assertEqual(self.svc.with_shared({}, "anthropic")["output_config"], {"effort": "low"})
+
+    def test_off_again(self):
+        self.req("/settings", {"defaults": {"temperature": 0.3}})
+        code, b = self.req("/settings", {"defaults": None})
+        self.assertEqual((code, b["shared"]), (200, False))
+        self.assertFalse(os.path.exists(self.svc.shared_path))
+        self.chat()
+        self.assertNotIn("temperature", self.engine.last_sampling)
+
+    def test_only_strata_s_own_page_may_set_them(self):
+        code, _ = self.req("/settings", None, {"Content-Type": "text/plain"}, raw=b'{"defaults": {"temperature": 1}}')
+        self.assertEqual(code, 415)
+        code, _ = self.req("/settings", {"defaults": {"temperature": 1}}, {"Origin": "http://evil.example"})
+        self.assertEqual(code, 403)
+        code, b = self.req("/settings", {"defaults": {"temperature": 9}})
+        self.assertEqual(code, 400)
+        self.assertIn("temperature", b["error"]["message"])
+        self.assertEqual(self.svc.shared, {})
+        host = self.base.split("://", 1)[1]
+        code, _ = self.req("/settings", {"defaults": {"temperature": 1}}, {"Origin": "http://" + host})
+        self.assertEqual(code, 200)
+
+    def test_they_need_the_key_when_one_is_set(self):
+        self.svc.api_key = "secret"
+        try:
+            self.assertEqual(self.req("/settings", {"defaults": {"temperature": 1}})[0], 401)
+            self.assertEqual(self.req("/settings", {"defaults": {"temperature": 1}},
+                                      {"Authorization": "Bearer secret"})[0], 200)
+        finally:
+            self.svc.api_key = ""
+
+
+class WebApp(unittest.TestCase):
+    """The web app (PR #22's dashboard idea, rebuilt): its page and files, and GET /metrics."""
+
+    @classmethod
+    def setUpClass(cls):
+        tok = ByteTokenizer()
+        cls.svc = Service(RecordingEngine(tok, "</think>\n\nhello", max_context=CTX), tok,
+                          ChatTemplate(ROOT / "serve/chat_template.jinja"))
+        cls.httpd = serve(cls.svc, port=0)
+        cls.base = f"http://127.0.0.1:{cls.httpd.server_address[1]}"
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.httpd.shutdown()
+
+    def get(self, path, headers=None):
+        req = urllib.request.Request(self.base + path, headers=headers or {})
+        try:
+            with urllib.request.urlopen(req, timeout=10) as r:
+                return r.status, r.headers.get("Content-Type", ""), r.read()
+        except urllib.error.HTTPError as e:
+            return e.code, e.headers.get("Content-Type", ""), e.read()
+
+    def test_page_and_files(self):
+        code, ctype, body = self.get("/")
+        self.assertEqual(code, 200)
+        self.assertIn("text/html", ctype)
+        self.assertIn(b"/web/app.js", body)
+        for path, want in (("/web/app.js", "javascript"), ("/web/app.css", "text/css"), ("/web/tokens.css", "text/css"),
+                           ("/web/components.css", "text/css"), ("/web/sprite.svg", "image/svg+xml")):
+            with self.subTest(path=path):
+                code, ctype, _ = self.get(path)
+                self.assertEqual(code, 200)
+                self.assertIn(want, ctype)
+
+    def test_only_the_app_files_are_served(self):
+        for path in ("/web/..%2Fserver.py", "/web/index.html", "/web/test.py", "/fonts/..%2F..%2Fsetup.py",
+                     "/fonts/missing.woff2", "/fonts/x.ttf"):
+            with self.subTest(path=path):
+                self.assertEqual(self.get(path)[0], 404)
+
+    def test_metrics(self):
+        data = json.dumps({"model": "m", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 5}).encode()
+        urllib.request.urlopen(urllib.request.Request(self.base + "/v1/chat/completions", data=data,
+                                                      headers={"Content-Type": "application/json"}), timeout=10).read()
+        code, ctype, body = self.get("/metrics")
+        self.assertEqual(code, 200)
+        m = json.loads(body)
+        for key in ("engine", "live", "requests", "hardware", "hardware_static", "history"):
+            self.assertIn(key, m)
+        self.assertEqual(m["engine"]["max_context"], CTX)
+        self.assertEqual(m["live"]["state"], "idle")
+        self.assertEqual(m["requests"][0]["output_tokens"], 5)
+
+    def test_metrics_need_the_key_when_one_is_set(self):
+        self.svc.api_key = "secret"
+        try:
+            self.assertEqual(self.get("/metrics")[0], 401)
+            self.assertEqual(self.get("/metrics", {"Authorization": "Bearer secret"})[0], 200)
+            self.assertEqual(self.get("/")[0], 200)                  # the page itself asks for the key
+        finally:
+            self.svc.api_key = ""
+
+
+if __name__ == "__main__":
+    unittest.main()