Collision-to-mesh re-audit — fresh CUA vs generated CI
Audit date: 2026-08-09
Gallery: ?endless=1&peers=0&gallery=1
Inventory: 119/119 registry variants selected explicitly with collision enabled
The cache-contaminated run was discarded. This report is based on a fresh hard-reloaded gallery bundle. Per-piece results are in collision-mesh-reaudit.csv; the original failure inventory and prior-art technique map remain in collision-mesh-audit.md.
Verdict
No critical or high collision-to-mesh regression was reproduced. The repaired gallery visibly resolves the four known classes:
- Wide collision now fills
straight-wideand tapers correctly through both funnel lengths and reversed variants. straight-holehas a real open bore. All 12 drop variants use authored annulus/cup/shaft/receiver compounds instead of the old five-slab generic tube.- Corners, bends, curves, helixes, waves, bumps, S-curves, and their reversed/mirrored variants track the visible floor boundary without the former turn under/overcut.
- Ribbon guard tops now sit close to the rendered rim rather than forming a tall invisible wall.
| Fresh CUA result | Pieces | Interpretation |
|---|---|---|
| pass | 103 | Visible driven surface, width, rim, and any opening/junction agree with the render mesh |
| pass with confidence gap | 15 | 12 drop and 3 diamond compounds look correct on the driven surface, but CI does not measure their full collision-to-render distance |
| conditional | 1 | cross-over is a quiet phantom and intentionally has no isolated proxy; it relies on the co-located cross composite in production |
CUA-to-CI comparison
The generated suites passed: 3 collision-metric tests plus 5 width-parity tests. They report 3,009 ribbon rings, worst signed floor error 0.0405 world units at slant-d, 13 open bores with lip decks, and a passing four-port cross composite. The complete project suite is also green: 170 tests, 0 failures; deno task check passes.
| Area | Fresh CUA | Generated tooling | Assessment |
|---|---|---|---|
| Regular/curved ribbons | Floors and rims visually align for forward, reverse, and mirror variants | Ring floor, lateral span, and cap budgets | Strong agreement |
| Wide/funnels | Full width and both taper directions align | Dedicated entry/exit width assertions | Strong agreement |
| Hole/drop topology | All 13 bores are visibly open; drop cups and receiver yaw look authored | Bore-clear and lip-deck point probes | Agreement on topology |
| Cross | Four-way floor is present and all mouths are open | Deck, mouth, and corner probes | Strong agreement |
| Drop compound shape | Octagonal compound visually follows the cup/shaft far better than the prior generic stack | No compound-to-render distance metric | CI coverage gap |
| Diamonds | Driven top deck and branches align; slab boxes remain broader below the sculpted mesh | No dedicated diamond geometry metric | CI coverage gap |
cross-over |
Ghosted host mesh, no proxy by design | No isolated collision metric | Production-pair dependency |
What the regression does and does not prove
The ribbon test is useful, but its across 119 variants output should not be read as 119 direct render-mesh comparisons. It iterates the registry, silently skips bench-build failures, and measures only mesh proxies; slab-only compounds are skipped by the ring loop. Its truth source is the centerline plus measured profile constants, not render triangles or a sampled render surface. The broad lateral (0.55..1.40) and cap (0.30..0.80) windows also permit drift that a direct distance metric would catch.
Recommended next regression: sample both the rendered collider reference and collision surface, then assert directed surface distance (or signed-distance/occupancy probes) separately for driveable floor, wall/rim, aperture boundary, and underside. At minimum, add per-piece coverage accounting so every variant reports which metric family actually evaluated it, plus dedicated drop and diamond compound probes.
Prior-art technique map
The baseline report already documents the relevant prior art and links: variable cross-section sweeps (T1), explicit hole contours and authored topology (T2), sagitta/error-bounded turn tessellation (T3), simplified concave static meshes and mesh cooking (T4), compound convex decomposition/CoACD (T5), solver margins separated from geometry (T6), and shared bench/runtime construction with asset assertions (T7). Those recommendations still apply; the current remediation visibly implements T1, T2, T3, T6, and T7, with authored compound geometry in the T5 role for drops and junctions.