Record actual 100k lexical index performance on Windows CPU

docs/SCALE-PERFORMANCE.md

@@ -1,6 +1,7 @@
 # Large-index lexical performance protocol
 
-Status: **1,000-record smoke executed; large run deferred**. No production
+Status: **1,000-record smoke and 100,000-record intermediate run executed;
+million-record run deferred**. No production
 release gate is passed by this document.
 
 `tools/scale-eval` uses the actual shared Rust `SearchIndex` engine, with no
@@ -119,3 +120,35 @@ tests passed in normal and optimized Python modes; formatting and strict
 release all-target Clippy passed. A subsequent Clippy-only parity-expression
 cleanup did not change the recorded smoke executable. The million-record run,
 representative libraries and UI/semantic timings remain open.
+
+## Windows VM intermediate scale, 2026-10-13
+
+Actual release diagnostic at source `59908ad`, clean worktree, used 100,000
+file records, 10,000 short text documents and 30 requests per case per mode.
+All assertions and integrity checks passed. Report:
+`.searchi/qa/scale-100k.json`, SHA-256
+`25a300f4912d33c07963eb712a9214879c326e8c1a16eeca130855798c561b71`.
+Executable SHA-256:
+`1c3a8774195ff13e2ffcbb4586446d1de68e2ac927452fd8c004ca4e7d09fd82`.
+Dependency lock is unchanged from the smoke. Rust 1.99.0, LLVM 23.1.1,
+Windows x64 MSVC; Q35 VM exposes Ryzen 7 7800X3D and eight logical CPUs,
+17,109,004,288 bytes RAM. Storage/power configuration was not independently
+measured. The installed desktop and its query worker remained running; no
+benchmark-owned inference was launched. Background activity was not controlled.
+
+| Case | Basic warm p95 ms | Debug warm p95 ms |
+| --- | ---: | ---: |
+| Exact name | 1.2085 | 1.4841 |
+| Unique keyword | 0.7019 | 1.1082 |
+| Common keyword | 9.9522 | 9.7389 |
+| Common filename | 0.6121 | 0.7390 |
+| Strict combined filters | 6.9529 | 7.2662 |
+| Absent query | 0.5178 | 0.8753 |
+
+Seed insertion took 28,224.5 ms and produced a 110,481,408-byte database.
+Reader opening took 3.83 ms; first unprepared input 78.02 ms; production prewarm
+283.56 ms. Reopening took 3.99 ms without flushing storage caches. A live
+process observation reported 15,540,224-byte peak working set at that moment;
+it is not a complete lifetime peak-memory measurement. This intermediate
+synthetic result does not close the million-record, real-library, concurrent
+indexing, semantic or UI gates and does not alter product ranking.