research: startup overlap + parallel route validation (L107, not merged)

Exact (24/24 hashes) but marginal: base compile ~16 ms shorter; whole view -1.6..2.2% at W8, near-ready mixed (TN W8 +31 ms, VS W12 +11 ms). Kept as patches/startup-overlap-not-merged.patch per the owner's simplicity rule (only merge measured wins on near readiness). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-28 22:50:23 +01:00
1 parent 20f1b35967
commit ef22b03851
@@ -0,0 +1,47 @@
+# Linux 107: base-shape startup
+
+## Result
+
+Kept the exact startup overlap and parallel route check. Across five interleaved W8 pairs, whole-view readiness improved by a paired median 35–55 ms (1.6–2.2%) in all four strates. W12 was mostly flat outside SW. No passage support rules or density predicates changed.
+
+## Critical path
+
+In FI W8, base compile ran from 59.407 to 224.529 ms. LOD0–1 admission followed at 224.580 ms; the first tile was ready at 246 ms, LOD0–1 drained at 1.302 s, and LOD2 completed at 1.397 s. Base shape remains a serial prefix before LOD0–1 work.
+
+Route validation does not supply tile data; it is a hard compile/publication gate. Its exact density samples now run in parallel, then failures are reduced in canonical route, segment, and sample order. It still gates compile completion. The surface-distance band is already parallel and is read by exact density queries for candidate nodes, so this round leaves it eager. Owner-shape extraction and sign-edge repair were already parallel. Packed intervals now overlap X/Y candidate sweeps and passage prep with split worker budgets and disjoint outputs.
+
+## Substage snapshots
+
+FI/TN W8/W12 profile traces, L106 → L107, milliseconds:
+
+| Stage | L106 | L107 |
+|---|---:|---:|
+| Owner shapes | 21.8–23.2 | 22.3–25.1 |
+| Packed + passage prep + candidates on path | 56.5–59.5 sequential | 45.3–54.5 overlapped |
+| Detail band | 49.7–83.3 | 54.2–77.5 |
+| Sign-edge repair | 9.4–10.3 | 9.7–11.6 |
+| Route validation | 25.7–26.9 | 18.4–18.6 |
+| Base compile | 173–205 | 157–189 |
+
+## Readiness
+
+Each value is the median of five complete runs. Parentheses give the median of the five matched after-minus-before pairs, in milliseconds. Negative is faster.
+
+| Stratum | W | Near LOD0–2 median (s) | Whole-view median (s) |
+|---|---:|---:|---:|
+| SW | 8 | 0.757 → 0.752 (−9) | 1.927 → 1.867 (−42) |
+| SW | 12 | 0.769 → 0.736 (−43) | 1.810 → 1.771 (−39) |
+| FI | 8 | 1.553 → 1.537 (−10) | 2.551 → 2.565 (−55) |
+| FI | 12 | 1.436 → 1.443 (+6) | 2.393 → 2.389 (−6) |
+| TN | 8 | 1.461 → 1.492 (+31) | 2.396 → 2.425 (−39) |
+| TN | 12 | 1.314 → 1.327 (+18) | 2.174 → 2.192 (+19) |
+| VS | 8 | 1.046 → 1.049 (−13) | 2.018 → 1.971 (−35) |
+| VS | 12 | 0.877 → 0.952 (+11) | 1.735 → 1.827 (−2) |
+
+## Checks
+
+- Exact hashes: 24/24 cases match L106 per-tile and whole-view hashes (SW/FI/TN/VS × Clang/GCC × W1/W8/W12); zero mismatches.
+- Fast tier: `SUMMARY PASS` (269 pass, five existing diagnostics).
+- Linux UE build: succeeded in 68.9 s with `-NoUBA -MaxParallelActions=4`.
+- UE arrivals: 3 FI and 3 SW succeeded, with no world-test failures. Unreal ran with taskset and timeout, without `prlimit`.
+- Logs and machine-readable results: `Saved/Linux/l107/`. Candidate source stamp: `5386228461b81b97f5649f6dc4beab5e3b9b8fe38ea706064a825d3ac77e65d3`.