chore(release): 0.0.1b2 — the beta corrected (#164)

[JArmandoAnaya]

Closes #164. The 0.0.1-beta.2 close-out: version bump, changelog, and the exit criteria checked against the artifact.

The milestone

0.0.1b1 shipped with every gate green and three defects a manual pass over the built wheel found within an hour. Five issues fixed, in the order the epic froze:

#158	the export report said polygons were not carried while YOLO and VOC wrote them as boxes	#165	4523ccb
#160	the annotator was unreachable from the UI	#166	5b390c4
#159	the gallery rendered one tile per row at every width	#167	54e4cdc
#162	the schema editor's swatch contradicted its own dot	#168	66c19ff
#161	/favicon.ico 404'd — the only console error in a clean load	#169	be09e25

All three major defects were the same shape: a claim verified against itself rather than against the artifact. So each fix landed with the guard that was missing, and each guard was verified by mutation — reintroducing the defect turns a named test red, checked one at a time.

Standing constraints, all held

FORMAT_VERSION	12 throughout. No migration added by any task — one would have meant a fix that went deeper than a fix.
openapi.json	moved once for a contract change, in #158, and once here for info.version. The generated client is byte-identical after this bump, which docs/releasing.md already documents as correct rather than suspicious.
the three gate words	confirm=, allow_destructive=, allow_lossy= still separate; #158 added a fourth state to the report and moved none of them.
kernel purity / annotator boundary	lint-imports 2 contracts kept, 0 broken.
one runnable example per milestone	none added — this milestone adds no capability.

What this commit does

VERSION → 0.0.1b2, carried to the three frontend packages as 0.0.1-beta.2 by pnpm version:sync. tests/test_versioning.py's pin moves with it, which is what makes a bump a deliberate edit in two places. CHANGELOG.md becomes a released section; CONTRIBUTING.md's version table gains a row; docs/releasing.md's runbook shows the version being cut, while the #69 argument above it stays historical.

Exit criterion 3 — the artifact, not the tree

• bash scripts/build_dist.sh → visionset-0.0.1b2-py3-none-any.whl (589 KB), with _static/index.html and the new _static/favicon.svg inside it.
• VISIONSET_REQUIRE_WHEEL=1 uv run pytest tests/packaging — 11 passed.
• Installed into a clean venv; visionset --version → 0.0.1b2, _static/ carries 5 files.
• The whole browser cycle driven against that installation: token → project → schema → ingest → the draft batch's inert tiles → the grid re-flowing at two viewport widths → a tile click landing on the clicked asset → back to the gallery → three annotations → complete → promote → publish → verify → export → download → clean console. 1 passed.
• formats: the export report says polygons are not carried while YOLO and VOC write them as bounding boxes #158's own reproduction replayed through the installed distribution — 3 boxes, 2 lane polygons, 1 weather tag, exported to yolo, voc and coco, with the report compared against the rows actually on disk:

yolo  held=6 on-disk=5 excluded=1 degraded=2 -> AGREES
      lane        degraded  yolo writes a polygon as its bounding box; the shape is lost
      weather     dropped   yolo cannot place a classification_tag and drops it
voc   held=6 on-disk=5 excluded=1 degraded=2 -> AGREES
coco  held=6 on-disk=5 excluded=1 degraded=0 -> AGREES

The issue reported excluded_annotations: 3 and found four label rows where two were expected. It now reports 1 and writes 5 of 6.

Every gate, locally

uv run pytest	exit 0
uv run lint-imports	exit 0
uv run mypy src/visionset	exit 0
uv run ruff check / ruff format --check	exit 0
pnpm -r build / pnpm -r lint / pnpm -r typecheck	exit 0
pnpm test	711 annotator + 123 ui-core vitest, 25 node:test
pnpm --filter @visionset/app e2e	76 passed
pnpm --filter @visionset/app cycle	1 passed (source), 1 passed (installed wheel)

After the merge

v0.0.1-beta.2 is tagged on the merge commit. Publishing to PyPI needs credentials this repository does not hold — docs/releasing.md step 4 — so it stays a human step, as it was for 0.0.1b1.