diag: passage vs boundary-band ceiling split (L88); coarse passage cache rejected

Verified ablations (zero passage work / zero band calls). LOD0-2 @W12: exact FI 2.27 / TN 2.25 s;
passages off FI 0.81 / TN 0.85; bands off FI 2.11 / TN 2.06. Passages cost ~1.4-1.6 s of near
readiness, bands ~0.1-0.2 s. A tile-local 4-voxel trilinear passage-delta cache increased carve
segment visits (FI 34.5M -> 56.2M) and slowed LOD0-2 (FI 2.58 -> 3.77 s); 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/Linux88/REPORT.md

@@ -0,0 +1,35 @@
+# Linux 88 — passage and boundary cost
+
+## Ceiling split
+
+L80 protocol: one warmup, then three interleaved measured runs per mode, worker count and stratum. Values are median seconds to LOD0–2 ready / whole view ready.
+
+| Stratum | Workers | Exact | Passages off, bands on | Bands off, passages on |
+|---|---:|---:|---:|---:|
+| FI | 8 | 2.419 / 3.696 | 0.808 / 1.738 | 2.306 / 3.556 |
+| FI | 12 | 2.273 / 3.463 | 0.812 / 1.683 | 2.114 / 3.290 |
+| TN | 8 | 2.458 / 3.624 | 0.912 / 1.814 | 2.222 / 3.308 |
+| TN | 12 | 2.253 / 3.294 | 0.851 / 1.683 | 2.056 / 3.061 |
+
+Passage composition costs about 1.40–1.61 s of FI/TN LOD0–2 readiness. Boundary bands cost about 0.11–0.24 s. The split counters confirmed the ablations: passages-off had zero passage queries, iterations, carve/support visits and relief calls, while retaining 11.39M FI / 7.72M TN band calls in profiled tiles. Bands-off had zero band calls while passage iterations remained 14.06M FI / 12.43M TN. The profile covered 629 FI and 614 TN near tiles. Passage-off intentionally failed the route walkability check because the remaining bands close an authored mouth; a split-only diagnostic continued the timing run.
+
+Exact-field readiness across all four strates:
+
+| Stratum | W8 LOD0–2 / whole | W12 LOD0–2 / whole |
+|---|---:|---:|
+| SurfaceWorld | 1.093 / 2.340 | 1.024 / 2.157 |
+| FloatingIslands | 2.419 / 3.696 | 2.273 / 3.463 |
+| TunnelNetwork | 2.458 / 3.624 | 2.253 / 3.294 |
+| VerticalShafts | 1.312 / 2.386 | 1.128 / 2.080 |
+
+## Prototype and decision
+
+I tested a tile-local 4-voxel trilinear cache of passage-composition deltas, with exact evaluation near or across the zero surface. It did not remove enough route work: instrumented FI/TN runs recorded 4.75M / 7.71M exact coarse-node evaluations, and carve-segment visits increased from 34.5M / 33.1M to 56.2M / 76.5M.
+
+Two unprofiled interleaved W8 pairs confirmed the regression. FI LOD0–2 median rose from 2.583 s exact to 3.765 s coarse; TN rose from 2.646 s to 4.649 s. Whole-view medians rose from 3.957 to 8.708 s for FI and 3.735 to 7.976 s for TN. The prototype was removed from source.
+
+The pre-measurement acceptance bar was max displacement ≤5 cm, mean ≤1 cm per stratum, and zero sign flips on canonical walk-floor samples. Displacement was not measured. The prototype stopped at the speed gate, so post-prototype passage audits, captures, and Clang/GCC W1/W8/W12 comparisons were not run. No evaluator version bump or hash re-pins were made; version remains 214. The fast tier passed on the cleaned exact-field source: 274 checks in 311.3 s, stamp `81a934b80ed90d2ab0137345d15688a9bd201e4ec6ce7af94b08f4cb4e77c1a3`.
+
+## Runs
+
+All benchmark and verification jobs used CPUs 0–11; headless jobs were capped at 6 GiB. No run exceeded 30 minutes. The timing and per-tile work logs are under `Saved/Linux/l88/ceiling-split/` and `Saved/Linux/l88/coarse-prototype/`; the fast-tier log and summary are under `Saved/Linux/l88/logs/fast-tier.log` and `Saved/Linux/l88/verify-fast/fast-summary.json`. No Unreal `Source/` files changed, so no UE build was needed. No commit, push, or stash was made.