diag: passage-disabled ceiling (L86); readiness unchanged without passages

Environment-only diagnostic hook disables passage composition (never a world setting).
Cold LOD0-2 / whole @W12 WITHOUT passages: SW 1.12/2.25, FI 2.18/3.34, TN 2.15/3.15, VS 1.12/2.03 s
- the same as with passages (FI 2.16 s). Passages are ~58% of sampled CPU but not the
readiness bottleneck. A lazy exact passage-result cache was +47% slower (24% hit rate, ~700 MB)
and was removed. No world change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1

Standalone/Research/WorldAssemblerM3/Linux86/REPORT.md

@@ -0,0 +1,28 @@
+# Linux 86 — passage ceiling and cache falsifier
+
+## Passage-disabled ceiling
+
+L80 protocol: one warmup per stratum and worker count, followed by three interleaved W8/W12 pairs. Values are median seconds to LOD0–2 ready / whole view ready. The diagnostic removes route carving, support composition, transition-mouth passage fade, and per-tile passage feature lists.
+
+| Stratum | W8 | W12 |
+|---|---:|---:|
+| SurfaceWorld | 1.008 / 2.194 | 1.122 / 2.253 |
+| FloatingIslands | 2.343 / 3.589 | 2.179 / 3.337 |
+| TunnelNetwork | 2.343 / 3.451 | 2.153 / 3.151 |
+| VerticalShafts | 1.327 / 2.289 | 1.123 / 2.026 |
+
+The FI/TN ceiling remains more than two seconds. In representative W12 profiles without passage composition, marching-cubes grid fill used 15.7 worker-seconds for FI and 14.8 for TN across LOD0–2; skirts used another 0.72 worker-seconds in each. The mesh grid work is the main remaining cost. Shape compilation was about 0.24–0.25 s wall time.
+
+Logs and raw runs are in `Saved/Linux/l86/ceiling/`.
+
+## Prototype and decision
+
+I tried a sparse, lazy world-coordinate map keyed by the exact float coordinate and incoming density bits. A miss ran the existing full passage composer and stored its exact post-composition output; a repeat became a lookup. This preserves carve ordering, support restoration, deferred support, and floors without interpolation. Miss computation ran inside tile generation, so it overlapped generation and remained in cold readiness.
+
+One FI W8 falsification smoke took **3.353 / 5.621 s**, versus L85’s normal-field W8 medians of **2.286 / 3.469 s** (+47% / +62%). It recorded 4.67M hits and 14.57M misses (24.3% hits), about **699 MB** estimated map storage, and **9.996 worker-seconds** of miss computation. That CPU time ran inside tile generation; the measured 5.621 s whole-view readiness includes it. I stopped at this clear regression; the three-pair W8/W12 prototype matrix was not run. Smoke log: `Saved/Linux/l86/logs/cache-smoke-fi-w8.log`.
+
+The result cache prototype was removed from source. It used no interpolation, but max/mean displacement was not measured. Post-prototype support, capsule, grade, ripple, naturalness, mouth-profile, and three-interface audits; before/after captures; Clang/GCC peer determinism; and fast tier were not run after the speed gate failed. The exact-key design was not treated as a substitute for those gates. No hashes were repinned or version bumped.
+
+## Files and failed runs
+
+The only retained Core change is the L86 environment-only ceiling hook in `Standalone/Core/src/ShapeWorldInternal.h`; it is diagnostic and not a world setting. No `Source/` files changed, so no Unreal build was needed. The cache smoke exited successfully but failed the speed gate. No run exceeded 30 minutes. No commit, push, or stash was made.