feat: a suggested polygon follows the mask to a pixel tolerance

[JArmandoAnaya]

What changed

A suggested polygon now follows the segmenter's mask to within a pixel tolerance the person chooses, and its outline is smooth rather than a staircase.

The pipeline (src/visionset/inference/masks.py). The mask is traced along its pixels' edges rather than through pixel centres, so a lone pixel is its unit square and the ring describes where the mask ends. One pass of corner cutting over those unit edges turns the staircase a pixel grid imposes into a smooth line while moving no real corner by more than half a pixel — a 51 px square still comes back as four corners. That ring, reduced once at a quarter pixel, is the contour every answer carries. The polygon is then simplified at an absolute tolerance in asset pixels, and the promise is stated as a bound the tests measure: every point of the contour lies within tolerance of the polygon. On a 250 px-radius disc that is 49 vertices at 1 px and 600 at 0.5 px; the previous diagonal-relative setting gave about 25 at its finest.

The wire. SuggestRequest.tolerance replaces the three-step detail: default 1.0, refused outside 0.25–16 (never clamped, so applied.tolerance never reports a value the server did not use). parameters names "tolerance" for a polygon class and stays empty for a box class. detail is now an unknown field and is refused. openapi.json and the generated client are regenerated.

The editor. The annotator's simplifier mirrors the kernel's at the same absolute tolerance, held to it by the regenerated golden fixture at every stop from 0.25 to 16 px. [ doubles the tolerance and ] halves it along the power-of-two ladder; the slider runs on a doubling track between the two ends and a value it leaves between stops snaps to the nearest stop on the next bracket press. A coarse tolerance can leave a small object with nothing to show; the suggestions are kept, so stepping finer brings the shape back. The visible label stays "Detail".

Docs. docs/content/inference.md (the pipeline and the setting), docs/content/ui.md and docs/content/ui/annotator.md (the slider and the brackets).

Found, not fixed

• The canonical contour is larger than before (about 1 200 points for a 600 px-radius object). It is within the design's budget and travels once per click; no wire-size gate exists.
• A refine click made while a coarse tolerance has emptied the preview repaints the last shape for the duration of the request, which is the existing "keep the best answer anyone has" rule applied to a state that could not previously occur.

Test plan

Run locally on the branch head:

• uv run pytest tests/inference tests/kernel tests/server — 3303 passed, 30 skipped
• uv run lint-imports — 4 contracts kept; uv run mypy src/visionset — clean; uv run ruff check . / ruff format --check . — clean
• pnpm -r build && pnpm -r test && pnpm -r lint — annotator 1071, ui-core 1302, app 11, all green
• CI=1 … playwright test e2e/annotate.spec.ts — 89 passed
• pnpm generate:client:check — generated client matches openapi.json; regenerating the simplification fixture produces no diff

New coverage: the fidelity bound is asserted for a disc, a lobed blob and a rectangle at all seven tolerances; the edge trace is pinned on a single pixel, a diagonal pinch, an enclosed hole and a two-pixel piece; the fixture proves both mirrors agree point for point; Playwright walks the ladder from the keyboard and drags the slider to both ends.