Stage buffers: pinning is opt-in (STRATA_STAGE_PIN=1); on by default it corrupted IQ3_S decode after a long prompt in the release gate

docs/DETAILS.md

@@ -1443,8 +1443,9 @@ the pairs in which the switched arm won). They are here so you can try them on y
   returned: decode 4 slots, median tok/s B/A (pairs faster): RTX 3060 0.95 (0/6) with misses only, 0.91 (0/6) with the
   look-ahead; Tesla P100 0.97 (2/6) and 1.00 (4/6); 16 slots on the P100 0.96 (0/6); RTX 5070 0.98 (1/6). The reporter
   measured the same on 2x RTX 3090 (the misses per layer are below one there, so the CPU round trip stays). Left off.
-- **`STRATA_STAGE_PIN`: pinned stage buffers** (Zhong Uncle, #1237; **on by default** while the host has 3 GiB of RAM to
-  spare, `STRATA_STAGE_PIN=0` turns it off, `=1` pins without the RAM check). Where the experts are served from the GGUF in
+- **`STRATA_STAGE_PIN`: pinned stage buffers** (Zhong Uncle, #1237; **opt-in, `STRATA_STAGE_PIN=1`**. It was on by default in
+  development, but the 0.1.41 release gate found IQ3_S decode corrupted after a 4096-token prompt with it on, likely a stage
+  buffer recycled before its asynchronous copy finished; it stays off until that is fixed). Where the experts are served from the GGUF in
   place (`--mmap-experts`, or too little RAM for the arena) the buffers the cache fill copies from are page-locked, so the
   copy no longer goes through the driver's bounce buffer (about 0.4-0.6 GiB of pinned RAM, one buffer falls back to
   pageable when the driver refuses). Decode, median tok/s on, off, 6 pairs of whole runs: Tesla P100 with the RAM capped at

src/core/expert_source.cpp

@@ -1149,8 +1149,10 @@ bool FileExpertSource::copy_from_files(int64_t layer, int64_t expert, uint8_t* d
     return true;
 }
 
-// Pinned stage buffers (PR 1237): whether they are the default is decided by the A/B in notes-141-cuda.md.
-#define STAGE_PIN_DEFAULT 1
+// Pinned stage buffers (PR 1237): opt-in (STRATA_STAGE_PIN=1).  On by default they corrupted IQ3_S decode after a 4096-token
+// prompt in the 0.1.41 release gate (48 tokens of id 0; off = identical to 0.1.40.4): with page-locked buffers the cache-fill
+// copy is truly asynchronous, so a recycled stage buffer can be overwritten before its copy finishes.  Off until that is fixed.
+#define STAGE_PIN_DEFAULT 0
 constexpr uint64_t kStagePinFloor = 3ull << 30;   // RAM left available after a pinned stage buffer
 void FileExpertSource::StageBufFree::operator()(uint8_t* p) const noexcept {
     if (pinned) (void) cudaFreeHost(p);