Fr0zkaandClaude Opus 5.5
c091e4434f
The demo tunnel's hard floor cap used a vertical Z - floorZ on a sloped Bezier floor, which is not a signed distance, so marching cubes produced a staircase through the first 7 m (40-51 mm ripple, 82 mm steps). The cap now uses the grade-normalised signed distance with a floor-adjacent bore re-application fading over the entrance; mesh tile Z bounds include cavern bounds. New three-rail entrance audit: L46 fails, L47 passes. verify: the cost baseline exports only build inputs into Saved/Linux/<out>/baseline-<commit> and removes it; /tmp unchanged after a full tier. Demo off: owner hashes identical to L46. fast/medium/full PASS. Known: sawtooth at the tunnel wall foot (next round). Research R14: Windows/MSVC readiness review. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
4.3 KiB
Linux47 report
Entrance floor
The L46 mesh has a repeating staircase through the first 7 m: 39.9–51.4 mm ripple and 81.8–82.9 mm largest residual step. The cause was the final hard floor cap, applied after the mouth and terrain stamps, using vertical Z-floorZ on a sloped Bezier floor. That field is not a signed distance to the ramp, so Marching Cubes interpolated a distorted gradient; stamp fill beside the cap made the ridges visible. Voxel height quantisation was not the main cause.
The cap now uses the grade-normalised signed distance. The tunnel SDF is reapplied in a floor-adjacent band, with the wider correction fading over the first 10% of the entrance; the narrow floor correction continues along the ramp. This removes the ridges without changing the tunnel wall profile. Mesh tile Z bounds now include resolved cavern bounds too; without that, the exported ramp lost triangles past about 6.6 m.
The mesh audit samples 0–10 m at 0.125 m on three rails (−1.5, 0, +1.5 m), requiring ≥85% coverage, ≤15 mm ripple and ≤50 mm adjacent residual step. L46 fails all rails: coverage 0.63–0.68, ripple 39.9–51.4 mm, step 81.8–82.9 mm. L47 passes all rails: 100% coverage, ripple 0.4–1.2 mm, no measured residual step. The Core audit also passes 243 samples with 0.3 mm maximum step and 0.0 mm reported ripple. Buried L43 and segmented-ramp L45 controls still fail as expected.
Verifier, hashes and cost
The cost baseline now exports only Core, Source, Compat, and the build/helper scripts to Saved/Linux/l47/baseline-e016f6a20fcfac785c9870b3d1133fd5df47a6d4; an exit handler removes the export. It does not copy the repository to /tmp. /tmp was 28K before and 28K after the full tier; the saved top-level-entry comparison is identical and the baseline export was removed.
Source stamp / source manifest: 68328b0c2a154e261eba94a067eb7d879f37ef987a54758275b4d14e3e9746d3. Demo-off owner hashes match L46 exactly for both recipes, with W1=W12. Whole-view hashes (density/mesh/payload/topology): floating-islands 598e9daaa1744ccb / 3114e0f4f6dbf734 / c49c4e48bff74089 / cadd4edd10d4d8ef; surface-world 7dc75e2c09302258 / 0cb34c40ceec4ce3 / 9d23e8c46b515cf1 / 2962c9464991eedd. Demo-on determinism passed 30/30 cells.
Five-pair median cost: floating-islands 16.390→16.435 s (+0.27%); surface-world 1.722→1.717 s (−0.29%). The baseline build took 40.2 s.
Tiers, captures and failed runs
Fast PASS (195.9 s), medium PASS (567.8 s), full PASS (1606.5 s); zero failed checks. All used source stamp above. Final captures: ground entrance, looking in, overview, inside to dais, and low entrance floor.
Exploration failures: the first entrance gate caught a 171 mm step; stamp-only re-carving restored about 40 mm ripple and a 100 mm step; full-width re-carving failed the wall-shape check at t=0.58. The final mouth-local fade passed the wall check at t=0.64. The first fast invocation also stopped because the executable had not yet been built; the rebuilt fast tier passed. L46/L43/L45 negative controls are intentional failures and passed their expected-fail checks. Work stayed within the 15-trial budget.
Orchestrator review (Claude)
Accepted, with one new defect logged. The stamp matches the final Clang and GCC builds (68328b0c…); demo-off hashes
are identical to L46; /tmp is unchanged after the full tier (1.8 MB used). The entrance floor is smooth now, and
the cause, a vertical cap distance on a sloped floor, was measured before the fix.
New defect: a row of jagged sawtooth teeth along the wall foot, where the floor-adjacent band correction meets
the bore wall. It shows in entrance-looking-in.jpg and along the edge in the overview. The three floor rails (−1.5,
0, +1.5 m) don't reach the 2.6 m wall foot. This is the third time a fix has moved an artifact to where the
audit isn't looking (L46 inside-only, L47 rails only). L48 adds a whole-surface artifact audit.
R14: no likely MSVC or UHT blocker. The Unreal adapter does not resolve stamps or cavern graphs, so the L43–L47 demos are standalone-only until the adapter supplies those records. The legacy chunk callers agree with the micrometre layout only at the default 8 m width.