MEASURED COLLISION (audit round): the proxies now match the kit, and CI proves it

Full remediation of docs/collision-mesh-audit.md (119/119 pieces,
re-audited clean same day — docs/collision-mesh-reaudit.md), built on
headless GLB measurement (gltf-transform over marble-kit.glb, pattern
tools/measure_ports.ts). Measured truths, grid units: wall inner face
0.287, rim floor+0.25, floor bore ROUND r 0.25 in straight-hole and
BOTH end-hole flavors, wide inner wall 0.575, split-double walls at
branch ±0.287.

- PARITY (T7): segmentProxies() is the ONE channel builder for
  production, the ?gallery=1 bench, and the sims (the bench had been
  omitting wide scales — a diagnostic-only wiring defect that falsified
  five audit rows). gapSegmentsOf lives with it; rail.ts re-exports.
  Bench asserts finite mesh bounds, frames empty plots safely, and
  ghosts cross-over's host mesh (a deliberate noMesh phantom).
- APERTURES (T2/T5): PieceDef.holes → LoopSegment.holes opens a real
  bore in the ribbon (lip decks, side slivers, walls, under-deck guide
  ring). Drop-stacks get an authored measured compound: annulus deck,
  octagonal cup (r 1.0 world) with the entry arch open, octagonal
  shaft, receiver cup opened toward the true exit yaw, pan law kept.
- WIDTH (T1): widening is an additive wall SHIFT (0.576 world/side),
  not the old x2.43 section scale that put wide walls outside the
  visible +-1.7 mesh; the bench passes def-authored port widths so an
  isolated funnel tapers honestly.
- RIM (T6): CHANNEL_PROFILE tops at 0.62 over floor (measured rim 0.5
  + 0.12 solver cap), wall face 0.62 lateral; junction composite walls
  follow the same law (diamond/spur/braid tightened ~0.4/side to the
  measured split-double faces).
- TURNS (T3): arc stations bounded by outer-boundary chord sagitta
  (0.008 world) instead of a fixed angle; crease rounding stays as the
  documented kicker-prevention law.
- JUNCTIONS (T2): the cross cell emits ONE authored 4-way composite
  (deck + solid corner posts, all four port arches open) from the mesh
  visit; the phantom is silent — two same-height route ribbons used to
  wall each other's route off.

QUANTITATIVE CI (tests/collision_metrics_test.ts + collision_width_test):
ring floor/lateral/rim budgets (3009 rings, worst floor err 0.041 at
slant-d's intentional crest rounding), RENDER-derived directed distance
(40k GLB channel verts within 0.2 of the collision surface; corridor
excludes port-face bands — shared boundaries the neighbor's ring
covers — and drop shells' cosmetic corners), bore-openness probes (13),
drop compound probes (12), diamond compound probes (3), cross probes,
and per-variant coverage accounting (rings 113 / render-dist 101 /
probes 17 / paired-by-design 1; bench-build failures fail loudly).
The CI caught three real defects during bring-up: approach ribbons
overhanging drop bores (lip-trimmed), missing bore side slivers, and
thin cross corner Ls (now solid posts).

174 tests green, deno task check clean.

commit

38d562aafc37964742db8c997283d7e10eff4e7d

parent

d32bc7844122c0aa1810d61f2f1524fe6c9ba911

author

Dami <dami.etoile@gmail.com>