Fr0zkaandClaude Opus 5.5
4f93f3fffb
Native FI binds and its own route/clearance/mouth-support checks pass on seeds 0 and 10, but 14/16 seeds have no island arrangement that clears the fixed passage tubes and reaches both mouth heights; seed 10 has one composed collision; ripple 6-7 cm vs the 3.5 cm limit on both upper passages. Third failure of the 'islands adapt to fixed tunnels' premise (L133 x2, L140). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
5.4 KiB
L140 — Floating Islands planned around passages
Not accepted: 0/16 composed seeds passed. The new Core recipe reaches native S1 dispatch and meets its dedicated route, passage-clearance, and mouth-support checks on seeds 0 and 10. Seed 10 still has one blocked capsule in the composed owner route; both seeds fail the unchanged passage-ripple gate. Fourteen seeds cannot produce a corridor-aware plan.
Change
Core evaluator version is 239 and Floating Islands geometry version is 10. The owner adapter resolves the fixed passage layout first, then plans nine native islands and a walk route around the 0–1 arrival and 1–2 departure tubes. The plan records route and island anchors in the resolved layout. The clearance check reserves the 1.053 m passage influence reach plus the 0.880 m capsule half-height and checks five support probes at each mouth. Shape composition preserves native landing support and passage shell distance for this recipe. Direct legacy recipe fixtures retain their earlier attach-before-layout order.
S1 binds to FloatingIslands-native in the explicit Core owner mode. bUseVoxelForgeCore remains opt-in; the fast-tier default owner pose still reports the legacy analytic S1 path.
Composed sweep
The final-source Clang run covered seeds 0–15. Results and per-seed logs are in summary.json and Saved/Linux/l140/seeds/ (source stamp 6205e6c180ef32c41792dc3e0d8bf443a96df41974f40334af4bf3a157ab88da).
- Seed 0: the planner tried 11 detours; its 94.5 m route has 1.47 m conservative tube clearance. Native route audit: 128 points, 0 blocked capsule samples, 0 unsupported points, 0 grade failures. Passage check: 5,325 tube samples, 0 island conflicts, 676,275 route checks, 0 route conflicts, arrival and departure support 5/5, 2.824 m minimum excess clearance. The composed S1 route and seams pass. Ripple is 1/3: wall ripple is 6.85 cm on 0–1 and 7.10 cm on 1–2 against the unchanged 3.50 cm limit.
- Seed 10: the native route and passage checks pass, with 5/5 mouth support and 0 isolated route conflicts. The composed S1 capsule audit finds one blocked sample at
(-6.246, -28.615, -737.982) m, inside the resolved 0–1 passage; composed density is-0.002062while native island density is+1.730469. Ripple is again 1/3 (6.22 cm on 0–1 and 7.19 cm on 1–2). Seams and negative controls pass. - Seeds 1, 2, 7, 8, 9, 12, 13, 15: no candidate clears the tubes; best conservative clearances range from −5.994 m to −0.406 m. Seeds 3, 5, 6, 14: candidates clear the tubes but the supported island profile cannot reach both fixed mouth heights. Seeds 4 and 11: candidate islands escape the sealed vertical layout.
The dedicated gate is therefore 0/16. No geometry variant was attempted after these results. The repeated 0–1/1–2 ripple failure is also consistent with L67's closed authored-radius experiment: it measured 7.27 cm on 1–2 against the same 3.50 cm limit (L67 report).
Other checks
- Fast tier: PASS, 287.8 s; negative controls and existing worker-hash checks passed. It continues to verify the default legacy S1 dispatch.
- Determinism: PASS, six runs (Clang/GCC × W1/W8/W12). All produced the same 512-tile density, mesh, payload, topology, material, and exact-query hashes, with layout identity
resolved-layout-v3-fnv64:08da20d55ff6820f. Native recipe hash:0b6a23bc73a3b867. See the baseline andsummary.json. - Captures: ten before/after JPEGs passed the existing image-quality gate. The manifest records eye-height arrival, island edge, route between islands, departure, and wide silhouette views for both versions. See before, before edge, before route, before departure, before wide; after, after edge, after route, after departure, after wide.
- Linux VFHostEditor: PASS in 73.58 s. Build log:
ubt.log. - L80 timing: not completed. The machine gate passed (load 0.97, 2-second CPU idle 98.5%), but the candidate warm-up could not construct native S1 around the checked-in Unreal-exported passage layout: 0 detours, best clearance −3.509 m. No measured L80 pairs were recorded; the ±1% A/A floor remains the L137 reference. A separate 30-second setup probe with a regenerated seed-0 owner plan timed out during passage-route planning, so I did not run a mismatched or incomplete timing matrix.
git diff --check: PASS.
Remaining issues
The fixed exported owner layout and many generated seeds leave no route-clear island arrangement. On seed 10, the native-only clearance check also misses a collision that the final composed owner route catches. Ripple remains above the existing limit on both upper passages. Native S1 is still opt-in and is not promoted by this round.
Changed Unreal source: VoxelForgeCoreAdapter.cpp. Windows compilation remains for the owner.