BitOps final Mono-cache experiment, Windows x64 Unity6000.6.7f1 Editor Mono, 2026-11-12.
Baseline: BitOps at main af80061e390b3d4aa3c6cc4a0dd63eed20d4b056. Candidate uses de Bruijn Log2 uint/ulong lookups and constant ReadOnlySpan byte properties. ENABLE_MONO selects fixed array caches initialized once from those spans. Native DLL IL proves the three Mono cache fields, single cached-array load per lookup and constant RVA span getters. Caches allocate 160 table bytes plus three array headers at initialization, not per call. No IL2CPP or other-runtime performance claim follows.
Each of six fresh Unity processes passes all619cases, zero skips/failures, including609 existing cases and four exhaustive16bit-window cases plus six timing cases. Every input has an independent loop-reference result; each timed batch checks its checksum. Arrays4096 values, seed1170 Pcg shuffle, varied bit widths or counts1..40;64warm calls,>=20ms calibrated batches,21rawrounds. NUnit test-case output retains BIT-ROWS with all raw samples and loaded runtime MVID. Ratios include loop/checksum overhead and exclude preparation. The wide varied baseline varies2.306..9.335ns, so inspect every pair instead of treating the largest ratio as a guarantee.
The controller binds unchanged source hashes within each capture; both stdout/stderr EOF and nativeexit0 are required. Baseline MVID a486838b-2f8e-41eb-88f3-013ff9996424 and candidate5038f12e-14a6-4c6b-a0a0-af10ad769fcd are stable within their respective three runs and disjoint. Actual assembly hashes bind the offline IL records to the measured candidate.
REPLAY: use an isolated checkout of main af80061e, overlay overlay/ onto the repository, place input/ contents under progress/.artifacts/replay-bitops, then run native PowerShell bitops-paired-mono-cache-final.ps1 there. Its repository offset expects this hidden path layout; adjust owned temp-project folder and Unity editor path/version for your installation. The controller restores the candidate in finally. Other session changes outside these fixtures are not needed; the overlay retains all runtime changes that accompany the measured assembly. Generate missing metas using scripts/generate-meta.sh if needed. Run the Performance fixture explicitly; normal test gates exclude it. Never point this controller at a personal Unity project. Numeric timings and module IDs can differ on replay.
Constant-span-only and getter-inlining experiments were rejected before this final candidate because they regressed this Mono runtime. They are not shipped. All accepted numbers in bitops-mono-cache-summary.json come only from the final six captures.
