feat(reconstruction): TRELLIS.2 pipeline, welded watertightness, absent-not-zero confidence

[Mizunandayo]

Two honesty fixes that improve the InstantMesh pipeline now, and a TRELLIS.2
pipeline that is complete but not deployed.

Both fixes are the same bug: "not measured" reported as "worst possible"

Watertightness was measured on split vertices. UV mapping duplicates a vertex at
every chart seam. The surface it describes is unchanged, but is_watertight returns
false — so a textured mesh lost 30% of the confidence weight to a rendering
artefact. measure() now welds a copy before testing watertightness and volume.
A test builds a geometrically identical seam-split sphere and asserts it still
reports closed.

An absent measurement scored zero. field_decisiveness reads an empty deviation
array as 0.0, at weight 0.45 — most of a 39.6% score on a good mesh. A pipeline that
genuinely has no density field now has the component dropped and the weights
rescaled to 1.0, with a field-decisiveness-unavailable notice, exactly as
foregroundQuality is already handled.

TRELLIS.2 — complete, not deployed

MIT, 4B parameters, sparse-voxel, PBR materials and a real UV texture. Adapter,
Dockerfile.trellis2, contract enum, config and wiring are all in and gated.

Not deployed. Three builds failed: a wrong extension path, then a worker OOM at
MAX_JOBS=8 on an 8GB builder — Cloud Build reports that as INTERNAL_ERROR, which
reads as infrastructure rather than sizing. Fixed to MAX_JOBS=2 and never re-run.

A separate Dockerfile on purpose: it needs CUDA 12.4, a compiler and five extensions
built from source, and folding that into the working image would put TripoSR and
InstantMesh — both serving — behind a build that has never succeeded.

Also built, default off

• CPU texture bake (xatlas + numpy rasterisation, colour sampled at texel
  resolution). Works — 53× the colour detail — but UV unwrapping drops watertightness.
  Wrong trade for a physics system: mass comes from volume, volume needs a closed
  surface. TEXTURE_RESOLUTION=0, and the default is what matters, because
  deploy-all.ps1 uses --set-env-vars and replaces the whole environment.
• Six real photographs into the sparse-view stage. The rig is azimuths
  30/90/150/210/270/330 at alternating +20°/−10°. Observation beats invention, but the
  capture UX asks for four, so it never fires.
• Gemini's label as a Zero123++ prompt. The text pathway exists and is encoded,
  but both models were fine-tuned on an empty prompt.

Gates

127 reconstruction tests · ruff · mypy 25 files · no contract drift