docs(research): step 2 verdict -- exact per-sample density cannot reach 600 ms

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ

RESEARCH-SPEED-2026-09.md

@@ -35,6 +35,7 @@ The steps below describe **future plugin experiments**. This research pass makes
    **Step 1 results (2026-09-30, commit `41b001b`):** core cook **1.50 ms p50 / 2.28 ms p95** (passes); through RMC's collision route the same support bodies were ready **8.5–10.8 s** after their meshes existed (fails): the wait is RMC's path, not Chaos, so player-support collision must bypass it. CMC with a field override: **883 false-support frames in 100 underfoot digs** (fails): step 4 is **field-aware Mover**. Uploads: cap 16 = **639 applies/s**, drain p95 2.22 ms (misses 2 ms by 0.22 ms), no render/RHI backlog, so tune the drain and **do not group yet**. The "8× per-sample gap" was a denominator artefact: 263 ns was export *wall* time with 15 workers on a shared lattice (≈3.4 µs CPU per sample); the real per-sample CPU cost is ~2–3.4 µs in both paths, so the 12–16× reduction target stands.
 2. **Exact-float region-evaluator experiment.** Use an isolated, editor-only explore-commandlet path inside the plugin because the live composer, strate manager, generator and SDF cache depend on Unreal types. Keep gameplay default-off and compare at the same lattice points and operator order. The reference is the chosen **new-world float definition**, not old-season geometry. **Pass:** zero 16-bit quantized-value disagreements and zero route/capsule regression across at least five seeds; ≤**600 ms** for the full 259³ cold ring on six workers before certificates. Record candidate visits, cache checks and worker CPU. Failure means local compilation did not remove enough work; investigate before rewriting numeric math.
 
+   **Step 2 verdict (2026-09-30, attempts 2a/2b/2c, commits `41a9265`, `fd9f809`, `43c11cc`):** all exact (zero Q16/sign disagreements over five seeds, reach-shrink control caught, same hash on 1 and 6 workers), none fast. Best (2c): setup once per 32³ chunk, **1.595 µs/sample vs 1.678 live, 7.9 s vs 7.1 s wall**. Binning cut candidate *visits* 21→16 per sample, but room/tunnel/tunnel-core *evaluations* did not fall (48 M / 5 M / 38 M for 17.4 M points): the cost is the geometry evaluation itself, not the search; noise is ~2.6%. **600 ms is out of reach while every lattice point must reproduce today's density exactly.** That exactness bar was a comparison tool, not a requirement: the owner accepts a new world version, so the next step may *redefine* the evaluator (for example coarser room/tunnel geometry with interpolation, or skipping bricks proven uniform), judged on shape tolerance, walkability and cross-machine determinism instead of equality with the old path.
    **Step 2 result (2026-09-30, attempt C; pass bar not met, full compiler untested):** five-seed Q16 comparison and two-run raw-float determinism pass. The deliberate room-bin reach reduction to 0.5 is caught (18,107 Q16 differences, 735 sign differences), with no saturated samples. On the 17,373,979-point cold ring, the live `GetDensityAt` p50 is **6.776 s**, **1.679 µs worker CPU/sample**, and **29.164 worker CPU seconds**; the editor-only 8³ SDF-bin variant is **19.543 s**, **2.879 µs worker CPU/sample**, and **50.023 worker CPU seconds**. Both measured variants observed six workers. Sampled visits fall from **21.19 to 16.38/sample** (about 23%), but wall time is 2.88× slower and the 600 ms target is missed. The variant still calls `GetDensityAt` for every lattice point; its bins cover rooms, room-floor joins and whole tunnel chains, while segment loops, route/cache decisions, passages, landing posts, disturbances and player edits remain in the canonical per-sample path. This is a **candidate-bin ablation, not the planned compiled region program**; it falsifies per-sample 3D bin lookup as a sufficient optimization but does not falsify hoisting region decisions or compiling the full ordered program. Step 2 remains open until that full program is measured; do not start step 3's certificate work against this slower baseline. Full report: `Saved/BuildHost/HostProject/Plugins/VoxelForge/Saved/TestHarness/step2_region_eval_20260930_c_region-eval/region_eval.json`.
 
    **Launch record:** attempt A's guarded UBT invocation logged `Result: Failed` (JSON writer and `ParallelFor` overload compile errors); it exited before launching a commandlet. Attempt B built successfully but `-regioneval=1` was ignored by Unreal argument parsing, so the ordinary explore commandlet exited successfully without the expected JSON. Attempt C used a bare `-regioneval`, rebuilt the editor module, and completed the intended experiment. No attempt timed out, crossed the 8 GiB private-memory limit, or produced a crash report; C peaked at 7,137,738,752 private bytes. The runner was corrected to detect UBT's logged failure even when its process exit code is zero.