research: one continuous tube through the mouth (L149, 0/16; kept as patch)

Mouth edges matched passage position, radius, tangent (dot 0.99999) and floor slope, but the cubic departure edge
still curls ~1 m back alongside the passage within 5 m; shell floor violations remain; composed sweeps 0/16.
Sixth mouth-join round (L144-L149): stop and take the design question to the owner.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
This commit is contained in:
Fr0zkaandClaude Opus 5.5 committed 2026-11-01 15:24:49 +01:00
1 parent f21aab775f
commit b8382eed84
14 files changed
+1374

No files matched your search

@@ -0,0 +1,32 @@
# L149 — one tunnel through the mouth
**L149 does not meet the bar.** The mouth anchors now match in position, tangent, radius and floor slope, but the departure spline still turns back alongside the passage, and the composed shell audit still finds floor violations. I stopped after the third geometry attempt.
## Changes and result
- Planner cave and shaft edges now start from the passage mouth floor, use the mouth radius and tangent, and carry the mouth floor slope. Their runtime SDF uses the matching flat-bottom tube profile.
- Owner route samples now follow the actual mouth-edge curve 2 m from the joint. The portal lead was shortened from 6 m to 1 m, and the passage-edge clearance exclusion was reduced from 10 m to 2 m.
- Evaluator version is v238. Added Clang W1/W12 generated owner-hash baselines for surface-world style 1 and floating-islands style 2.
- These were the three geometry attempts: match the mouth tube; move route samples onto the curve; shorten the lead and clearance exclusion. The final seed-2 shell gate still failed, so there was no fourth attempt.
The measured premise was: matching the mouth geometry and shortening the lead would make the first/last route edge continue the passage as one forward tube. It is false for the current cubic edge. The seed-2 departure edge reaches about **1.00 m behind the mouth by 5.07 m of travel**, despite an initial tangent dot of 1.0. The shell audit independently finds three floor violations at passage 2–3.
## Measurements and checks
- **Focused seed 2, W8, clip radius 3, max level 0, final source stamp `19dbe26a…`:** route audits for strata 2 and 3 pass. Shell audit: 85,162 floor samples, 3 violations; 5,589 wall samples, 0 violations; 8,430 interior samples, 0 violations. The interior capsule audit passes with 281,043 samples and 3/3 negative controls. Ripple, naturalness and mouth-profile summaries pass. The audit exits 10 after the shell failure, before emitting the tile-seam audit; seam status is therefore **not measured** in these failing runs.
- **Builds:** final Clang and GCC headless binaries and the Clang offscreen renderer built successfully with source stamp `19dbe26ab76a629baadc54bb0ab93eaa22d6f878e6946df2b71509c12c9524e4`.
- **Mouth geometry audit:** all three seed-2 mouth edges match their anchor position and radius; minimum tangent dot is 0.9999999999886, maximum floor-slope delta is 0.000475, and all 3/3 reversed-tangent negative controls are detected. The first-12-m forward-extension check fails on the departure edge. Across the 16-seed sets it passes 1/16.
- **Composed sweeps, W8, seeds 0–15, clip radius 3, max clip level 8, max level 0:** Tunnel Network **0/16**, Vertical Shafts **0/16**. Both sets have 0/16 passing shell and passage summaries. The interior-air capsule summary passes 6/16 in each set. Naturalness, mouth profile, ripple and the interior, ripple-gap and blocked-route negative controls pass 16/16. The route audit passes 2/16 Tunnel Network cases and 8/16 Vertical Shafts cases. The seam audit is not reached because the shell audit exits first.
- **Clang/GCC worker matrix, seed 2, W1/W8/W12:** all six runs emitted the same 12-line pipeline hash signature and resolved layout identity `resolved-layout-v3-fnv64:d980465950135a9d`. This focused matrix used `--render-distance-chunks=0` and max level 0; it checks the seed-2 view and layout identity, not a full multi-LOD world. The fast tier also passed its owner W1/W12 hash baselines and random-camera W1/W8/W12 checks.
- **Fast tier:** **FAIL**, 356.1 s. Existing owner passage checks, v238 hash baselines, worker checks and negative controls passed. The new four-strate owner-plan stage failed its tunnel-shell and passage-summary checks, arrival/departure chain, and strata-2/3 arrival/departure walkability checks.
- **Captures:** three matched-camera before/after comparisons, rendered at 1280×720 from mouth height. All six images pass capture quality. Before images use the L147/v237 executable baseline; L148's saved exclusion-gate build rejected seed 0 before producing a layout. Its accepted seed-2 layout identity matches L147. After images use the final v238 source stamp.
| Joint | Before | After |
| --- | --- | --- |
| Seed 0, arrival | [Before](captures/before-seed0-tunnel-network-route-edge-0.jpg) | [After](captures/after-seed0-tunnel-network-route-edge-0.jpg) |
| Seed 0, departure | [Before](captures/before-seed0-tunnel-network-departure-edge.jpg) | [After](captures/after-seed0-tunnel-network-departure-edge.jpg) |
| Seed 2, shaft arrival | [Before](captures/before-seed2-vertical-shafts-route-edge-0.jpg) | [After](captures/after-seed2-vertical-shafts-route-edge-0.jpg) |
## Remaining risks
The v238 geometry still fails the requested composed shell and route gates, so it is not a passing L149 result. The captures are ready for the owner’s visual review. No Unreal module `Source/` files changed; `Standalone/Source/WorldAssemblyPrototype.h` did change.