test(geometry): add real corpus models as Geometry_IT fixtures

[jcschaff]

Extends the geometry golden suite (#2028) with real stored models, and wires the group into
regression CI.

Built the documented way: fixtures and goldens were generated on test/geometry-goldens-base
against the pre-change implementation (f35beaddcd), then cherry-picked here — so the goldens
provably predate the code they check.

Why real models, when synthetic fixtures already pass

The synthetic set is spheres, shells and stripes. Real segmentations are not shaped like that.

fixture	pixels	shape
corpus_209284198_600x300x22	3,960,000	3D, wide and thin
corpus_26454463_564x160x31	2,797,440	3D
corpus_95707047_208x153x83	2,641,392	3D, thick z
corpus_65311813_256x256x34	2,228,224	3D, cubic
corpus_12522025_2151x504_2d	1,084,104	2D, extreme aspect
corpus_201022999_211x201x11	466,521	3D, small

corpus_95707047 alone yields six regions from two pixel classes — five separate cytosol bodies.
Nothing in the synthetic set produces that, and multi-region-per-pixel-class is exactly the case that
region linking and merging has to get right.

The one design decision worth reviewing

These rebuild surfaces rather than pinning what the document already carries.

A stored model has a <SurfaceDescription>; XmlReader applies it, so precomputeAll skips
updateAll() on parse and the geometry arrives with regions restored from the file rather than
computed (this is the correction recorded in #2025 §4.1). Pinning those would test the XML reader
instead of surface generation — the opposite of the point. fromCorpus() therefore calls
updateAll() to force a fresh RegionImage and SurfaceCollection.

Naming: they are not actually slow

You asked for these as slower fixtures, and I want to be straight that the label overstates it —
2.87 s for all six. They could sit in Fast and run on every push instead of only in the
regression lane, which would be strictly better protection. I kept them separate for headroom: they
depend on the corpus resources, and the obvious way to extend them is to add more models. Easy to
promote if you'd rather.

CI wiring — the part that would have silently done nothing

regression.yml drives its matrix from an explicit group table, so a tag that isn't listed is never
selected: the tests would have compiled and never run. Added Geometry_IT as a single shard.

Verified by executing the workflow's embedded matrix script directly rather than trusting the
edit — 16 entries across 7 groups with Geometry_IT present, and the YAML parses.

Verification

pre-merge base branch	synthetic 7/7 (0.42 s), corpus 7/7 (2.87 s)
on master (post #2026/#2027)	-Dgroups=Geometry_IT → 7/7 (3.87 s)
vcell-core Fast	594 run, 0 failures, 8 errors — the documented Poetry-environment ones

The middle row is the experiment repeated for real models: goldens generated by the old
implementation, matched exactly by current master.

Also in here

Golden regeneration is now split, so iterating on the synthetic fixtures doesn't silently leave the
corpus goldens stale or spend minutes rebuilding them — GeometrySurfaceGolden.main with no
argument does both, with fast does only the synthetic set. And the comparison logic is shared
between the two test classes rather than duplicated, including the everyFixtureHasAGolden guard,
since a @TestFactory whose resources went missing would otherwise just produce fewer tests and
stay green.

Refs #2021, #2025, #2028

🤖 Generated with Claude Code

https://claude.ai/code/session_018kr8SbzXtwW3gMVUgMfDDt