wip: cave planner squash-merged into wip/combined (L173, RED - not for main)

Five conflicts resolved (CURRENT-STATE, ResolvedLayoutInternal.h, StreamBenchmarkHeadless.cpp,
StreamBenchmarkPipeline.h, WorldAssemblyPrototype.h) keeping liminal + cave planner; L156 band fit only; evaluator v239
with v239 baselines (18/18 Clang/GCC W1/W8/W12). Fast tier FAILS (FI upper arrival: 40 unsupported chamber samples,
10 chamber-floor shell violations); TN and VS composed sweeps 0/16 (passage, seam, shell gates). Living plane loads at
PostConfigInit (UE smoke passes). Source/: VoxelForgeCoreAdapter.cpp; VoxelForge.uplugin.

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

Standalone/Research/WorldAssemblerM3/Linux135/REPORT.md

@@ -0,0 +1,32 @@
+# Linux135 — Tunnel Network planner
+
+Source v236 (`8dc4294ce0250a591ee4515471c3c547e796ffa4346ab44f0ae1740bbc3797f5`) on `wip/l134-cave-planner`.
+
+## Route audits
+
+Seeds 0–15 pass the composed passage, seam, shell, naturalness, mouth-profile, ripple, and blocked-route controls (16/16). The negative control still detects all three blocked routes. No audit threshold changed.
+
+The mouth joint now takes its floor/profile from the resolved mouth geometry and carries that shared profile into the planner's first and last edges. This fixes seed 1's blocked arrival and seed 2's unsupported arrival. For seeds 0 and 3, the first 648 regular route candidates exceeded the grade bound; the bounded deterministic fallback finds a route. “No sweepable route candidates” was a planner search miss, not an empty cave.
+
+## Cost
+
+L80 medians are near-ready / whole-view percent deltas. The serialized 144-run matrix started at load 0.81 / 98.4% CPU idle. Each cell gives `assembler/world` first, then exact L134 (`a3c8e36`).
+
+| Layer | W8 | W12 |
+|---|---:|---:|
+| 0 | +3.88 / +0.27%; +1.79 / -0.21% | +8.22 / +4.09%; +3.89 / +2.49% |
+| 1 | +8.82 / +4.96%; +4.14 / +1.71% | +11.83 / +5.35%; +4.10 / +2.09% |
+| 2 | +13.33 / +3.77%; +2.34 / -0.09% | +12.56 / +5.20%; +1.40 / +0.61% |
+| 3 | +13.65 / +3.99%; -0.25 / +0.44% | +12.71 / +5.14%; +2.97 / +2.12% |
+
+The two speed changes were bounded cave AABBs with a 1 mm float guard, and normalized edge-BVH bounds with one normalization per sample. This improves on L134, but misses the `assembler/world` target: near-ready remains 3.88–13.65% slower. Layers 1–3 overlap the native cave volume and spend time on roughly 4.95–5.76 million near-LOD cave samples per view; layer 0 has none at LOD0–3 and is still up 3.88% at W8, showing some generic composer/classification cost as well.
+
+Planning is about 3.75–4.19 ms per view (cavern planning 0.46–0.49 ms). At LOD0–3, room/edge candidates average 0.39/0.23 per composer tile on layer 1, 0.47/0.27 on layer 2, and 0.06/0.04 on layer 3. Tile proofs skip work at LOD0–3; measured cave views have no tile skips at LOD4–8. The reported 180–320 ns density-call estimate includes all shape fields, not cave-only time. Profile: `Saved/Linux/l135/profile-summary-speed2.json`.
+
+## Gates and pictures
+
+Fast tier passes in 301.7 s. Clang and GCC each match 13 hash lines at W1/W8/W12. Route seams and the negative controls pass; canonical baselines are v236. The post-timing VFHost Linux Development build succeeds: 27/27 actions, 77.70 s, using `-NoUBA -MaxParallelActions=4`.
+
+The seven-view gallery has four planner views at the planned floor plus 1.6 m and three legacy views at legacy coordinates. Open `captures/COMPARE.html`; camera coordinates and renderer settings are in `captures/capture-index.json`.
+
+Sweep: `Saved/Linux/l135/optimized-seed-sweep.json`. Timing: `Saved/Linux/l135/timing/summary.json`. Determinism: `Saved/Linux/l135/determinism-summary.json`. Host build log: `Saved/Linux/l135/ubt-l135-console.log`.