Describe measurements instead of pointing at research outside this repo

Comments pointed at research notes, named by directory and file, that a
reader of this repository cannot open. The measurements are kept; only
the pointers go.

src/lib/src/direct.hpp

@@ -185,8 +185,8 @@ struct BlockData {
 // sliced by `asmop.size` (a SEPARATE ASM-side decode, `r_asm_disassemble`).
 // This producer calls only `r_anal_op` (never `r_asm_disassemble` -- that
 // call exists purely to build the mnemonic), reading `bytes` from its own
-// per-block/per-range read buffer sliced by `op.size` -- `instructions-api-
-// lab/FINDINGS.md` measured the two decoders never disagreeing on length or
+// per-block/per-range read buffer sliced by `op.size` -- measured, the two decoders never
+// disagreed on length or
 // on which addresses fail to decode, across five real/generated fixtures,
 // but this is a live risk to keep in mind if a future radare2 arch plugin
 // ever changes that.
@@ -284,7 +284,7 @@ struct FlagRow {
 // `RBinString` at all -- it is resolved per row via `r_bin_get_section_at`;
 // `vaddr` is rebased through `r_bin_get_vaddr`, not read off the raw member;
 // `type` is rendered through `r_bin_string_type`, not the raw char code.
-// Measured in a private lab at 26-86x across a real PE, a large generated
+// Measured at 26-86x across a real PE, a large generated
 // fixture, and a real large DLL, with exact field parity.
 struct StringRow {
   int64_t     addr = 0;
@@ -304,7 +304,7 @@ struct StringRow {
 // (`r_str_srwx_i`), not the raw permission bitmask, and `addr` silently
 // downgrades to NO rebase for a section lacking the READ permission bit --
 // a real per-row branch on the section's own `perm`, not a fixed `va` mode
-// for the whole table. Measured in a private lab at 19-58x (`sections`) and
+// for the whole table. Measured at 19-58x (`sections`) and
 // 58-195x (`segments`, vacuous on every PE fixture tested -- no PE section
 // sets `is_segment`, so this rides `sections`' proven-live shared producer)
 // with exact field parity.
@@ -326,7 +326,7 @@ struct SectionRow {
 // `RBinInfo::has_lit`. All of this goes through PUBLIC `R_API` calls, never
 // a private/static helper -- the producer calls the same functions the
 // emitter calls, it does not reimplement demangling or address resolution.
-// Measured in a private lab at 10-13x across a real PE, a large generated
+// Measured at 10-13x across a real PE, a large generated
 // fixture, and a real 1986-import DLL, with exact field parity (the
 // genuine C++-demangling branch itself is unexercised by any available
 // fixture -- radare2 is a C codebase, so every fixture's imports are
@@ -342,9 +342,8 @@ struct ImportRow {
 
 // One row of `comments` -- the command path's `CCj`, a single full-scan
 // command (not an N+1-per-function shape, unlike `funcs`/`xrefs`/`locals`/the
-// blocks cluster -- the same "one command already" shape `binary-meta-lab`
-// found for `strings`/`sections`/`imports`/`entries`). Measured in a private
-// lab at 5-70x with exact field parity, including two quirks found by
+// blocks cluster -- the same "one command already" shape measured for
+// `strings`/`sections`/`imports`/`entries`). Measured at 5-70x with exact field parity, including two quirks found by
 // reading `r_meta_print` (`libr/anal/meta.c`) rather than assumed from the
 // struct:
 //
@@ -376,7 +375,7 @@ struct CommentRow {
 // (exports), merged by address (this project's own `merge_row`, preserved
 // exactly on the API leg). `iej`'s JSON never emits `name`/`size`/`bind`
 // (confirmed by reading `bin_entry`'s JSON block field-by-field) but DOES
-// emit `type` via `r_bin_entry_type_string` -- an early lab probe omitted
+// emit `type` via `r_bin_entry_type_string` -- an early probe omitted
 // this and showed a real parity failure before the producer was corrected.
 // `entries.name` is NOT the demangled name despite living one call away from
 // a full demangling pipeline: reading `bin_symbols`' JSON emit block found
@@ -384,7 +383,7 @@ struct CommentRow {
 // name), UTF8-escaped -- the demangled form goes to a separate `"demname"`
 // key this table never reads. `its_an_export` (`!is_imported && bind ==
 // "GLOBAL"`) is a private/static helper in radare2 but trivial to reproduce
-// without calling it. Measured in a private lab at 9-78x across a real PE,
+// without calling it. Measured at 9-78x across a real PE,
 // a large generated fixture, and a real 755-export DLL, with exact
 // field parity.
 struct EntryRow {
@@ -625,7 +624,7 @@ class Api {
   // `fi->{addr,size,name,realname,space->name}` directly -- no `fs`
   // selection, no per-space re-listing, no session-state save/restore at
   // all, since `r_flag_foreach` is not scoped by the currently-selected
-  // flagspace the way the `fj` COMMAND is. Measured in a private lab at
+  // flagspace the way the `fj` COMMAND is. Measured at
   // 17-47x on a full scan across a real binary and generated fixtures, plus
   // a real large DLL, with exact field parity -- including the pointer-
   // equality `realname` quirk -- before being ported.
@@ -667,7 +666,7 @@ class Api {
   //
   // Unlike every READ producer above (an N+1-per-row scan, where command
   // dispatch is paid once per row), a single write is one command either
-  // way -- there is no N+1 shape to collapse. A private standalone lab
+  // way -- there is no N+1 shape to collapse. A benchmark
   // measured a BULK workload instead (N repeated writes in a loop,
   // simulating a bulk-rename/bulk-annotate script) and found each of the
   // four operations examined pays a real, per-call fixed cost for

src/lib/src/direct_libr.cpp

@@ -1085,7 +1085,7 @@ class LibrApi : public Api {
   // `func_addr` stays 0 on every row here). Refills its read buffer from IO
   // whenever fewer than 32 bytes remain -- the same low-water threshold the
   // command path's own `pdj` loop uses (`disasm.c`, `i >= nb_bytes - 32`,
-  // confirmed by `pd-range-lab`), since the longest x86 instruction is under
+  // confirmed by measurement), since the longest x86 instruction is under
   // 16 bytes and every other arch this project targets is shorter still.
   void instructions_range(int64_t start_addr, int64_t count,
                           std::vector<InsnRow>& out) override {

src/lib/src/tables_comments.cpp

@@ -29,7 +29,7 @@
 //
 // The WRITE path (this table's `text` UPDATE and INSERT) is ALSO now
 // C-API-backed on the libr flavor, a separate, later port from both of the
-// above: a private standalone lab measured a bulk workload (N=500 repeated
+// above: a benchmark measured a bulk workload (N=500 repeated
 // sets, simulating a bulk-annotate script) at 30-40x over the CMD leg
 // (`r_core_call_str_at("CCu base64:...")`, the SAME dispatch this table
 // already issues), with exact field/behavior parity once the API leg

src/lib/src/tables_flags.cpp

@@ -21,7 +21,7 @@
 // A full scan is that same enumeration: one `fj` (full dump) + one
 // `fs <name>; fj; fs *` PER FLAGSPACE to attribute the real space back onto
 // each flag by name. On the libr backend this reads `r_flag_foreach` directly
-// instead — measured in a private lab at 17-47x on a full scan across a real
+// instead — measured at 17-47x on a full scan across a real
 // binary, this project's own scale fixtures, and a real large DLL, with exact
 // field parity (including a pointer-equality quirk in radare2's own
 // `realname` field), before being ported (see `direct.hpp` for why the second
@@ -61,7 +61,7 @@
 // unbuilt and is not needed here.
 //
 // The rename is ALSO now C-API-backed on the libr flavor, a separate, later
-// port from the cmd_noeval rollout above: a private standalone lab
+// port from the cmd_noeval rollout above: a benchmark
 // measured a bulk workload (N up to 500 repeated renames, simulating a
 // bulk-rename script) at 48-62x over the CMD leg (the SAME
 // `r_core_call_str_at("fr ...")` dispatch this table already issues), with

src/lib/src/tables_funcs.cpp

@@ -20,8 +20,7 @@
 // pilot of this dispatch.
 //
 // `name`'s rename is explicitly REFUSED a C-API port, unlike the sibling
-// rename/set writes on `comments`/`flags`/`locals` (measured together in a
-// private standalone lab). Not a speed question — `afn`'s bulk-rename
+// rename/set writes on `comments`/`flags`/`locals` (measured together). Not a speed question — `afn`'s bulk-rename
 // cost clears the same 3x bar the others did. The refusal is PARITY: reading
 // `afn`'s actual handler (`__setFunctionName`, `cmd_anal.inc.c`) found it
 // does substantially more than `r_anal_function_rename` alone —
@@ -32,12 +31,12 @@
 //     reads back through `afcf`;
 //   * it fires the `on_fcn_rename` callback.
 // `r_anal_function_rename` touches ONLY the function's own name field and
-// hashtable entries — none of the above. The write-path lab proved this
-// live, not just from the source read: after an equal-sized bulk rename on
+// hashtable entries — none of the above. Measured live, not just
+// read from the source: after an equal-sized bulk rename on
 // two independent cores, the CMD leg's flag was renamed on 100% of rows and
 // the API leg's flag was renamed on 0% — a real, silent desync between
 // `funcs.name` and its own backing flag, the exact bug class already fixed
-// once this session for `flags.name` (`afn` creating a duplicate flag
+// for `flags.name` (`afn` creating a duplicate flag
 // instead of renaming in place). Porting this write would require
 // reimplementing `__setFunctionName` in full, which is not a "call the C
 // API instead" port at all — it is copying that problem in reverse. Stays

src/lib/src/tables_locals.cpp

@@ -13,7 +13,7 @@
 //
 // A full scan enumerates functions (`aflj`) and pulls each one's vars
 // (`afvj @ <addr>`). On the libr backend this reads `fcn->vars` directly
-// instead — measured in a private lab at 50-780x on a full scan across a real
+// instead — measured at 50-780x on a full scan across a real
 // binary and generated fixtures, with exact field parity, before being ported
 // (see `direct.hpp` for why the second producer lives in another TU). The
 // command path remains the only implementation r2pipe can have, and it also
@@ -26,13 +26,13 @@
 //
 // The WRITE side (this table's `name` rename and `type` retype) is ALSO now
 // C-API-backed on the libr flavor, a separate, later port from the read-side
-// scan above: a private standalone lab measured a bulk workload (N
+// scan above: a benchmark measured a bulk workload (N
 // repeated renames/retypes, simulating a bulk-retype script) at 78-150x over
 // the CMD leg. Reading `cmd_anal.inc.c`'s `afvn`/`afvt` handlers found each is
 // nothing more than `r_anal_function_get_var_byname` + `r_anal_var_rename`/
 // `r_anal_var_set_type` — the SAME calls the API leg makes directly, with no
 // other side effect (unlike `funcs.name`'s `afn`, refused for exactly this
-// reason on that table) — so parity is exact, proven live in the lab. The
+// reason on that table) — so parity is exact, proven live. The
 // single-function readback (`local_readback_matches`, below) stays
 // command-based on BOTH legs, same as `funcs`'/`flags`' own per-write
 // readbacks — this port only moves the WRITE call itself. On every non-libr

subprojects/libxsql.wrap

@@ -2,10 +2,9 @@
 # sqlite3/nlohmann). It ships no meson.build of its own (CMake-native),
 # so the packagefiles/ overlay below supplies one for the meson build path.
 #
-# In the monorepo, prefer the editable sibling checkout instead of this fetch by
-# pointing meson at it:  meson setup build-meson -Dlibxsql:... (or drop a
-# subprojects/libxsql/ checkout in place; meson uses a present directory over the
-# wrap). For a standalone/public build this fetches the same source CMake does.
+# When developing against a local libxsql, prefer that checkout over this fetch:
+# drop it in as subprojects/libxsql/ and meson will use the present directory in
+# preference to the wrap. Otherwise this fetches the same source CMake does.
 [wrap-git]
 url = https://github.com/0xeb/libxsql.git
 revision = main