research: passage preparation filters (L126, no-go, source restored)

Pair overlap comes from per-segment bounds expanded by radius + 4.335 m route reach; two exact filters cut
floor-profile calls 20-23% but near-ready stayed +3..+11% vs L109 (within ~2% of L125).

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

Standalone/Research/WorldAssemblerM3/Linux126/REPORT.md

@@ -0,0 +1,20 @@
+# Linux126 — passage preparation
+
+**No-go for the L109 speed bar.** Two exact-preserving preparation filters were tried. They reduced floor-profile calls by 20–23%, but did not reduce segment/sample pair counts or near-ready time enough. The tested source was restored after the host build.
+
+The L125 overlap is from each segment’s sweep bounds, expanded by its radius and the 4.335 m route reach. Adjacent segment bounds revisit the same samples. The feature slab only stamps outside sentinels; it does not create the pair count. Floor-profile and shell work multiply the cost of each pair. L125’s LOD0 pair overlap was 67.2%, 72.0%, 72.0%, and 67.6% for S0–S3. The first filter cut floor-profile calls from 3.107M→2.457M, 8.487M→6.549M, 8.303M→6.418M, and 5.269M→4.169M respectively.
+
+Five interleaved L80 measurements per cell, on a quiet machine (load 0.34, idle 98.2%):
+
+| Stratum | W8 near-ready vs L109 | W12 near-ready vs L109 |
+| --- | ---: | ---: |
+| S0 | +3.20% | +4.63% |
+| S1 | +9.34% | +7.03% |
+| S2 | +6.16% | +7.68% |
+| S3 | +10.97% | +8.65% |
+
+L126 was +0.52% to +2.13% vs L125 in six cells and within 0.20% in the other two. All eight cells missed L109.
+
+Quality stayed unchanged: 264 fast-tier checks passed; walkability passed 21,425/21,425 samples; the floor probe remained 7.884 mm; four seam pairs matched all 4,900 shared samples bit-for-bit; all 23 captures were verified and byte-identical to L125. Clang/GCC W1/W8/W12 hashes matched, and the v225 baselines were unchanged. The Linux VFHostEditor build succeeded in 37.60 s with `-NoUBA -MaxParallelActions=4` and no `prlimit`.
+
+No evaluator bump or baseline update was needed. Timing and count logs are under `Saved/Linux/l126/`; captures are under `captures/`.