fix: guard polygon construction in segmentation workers

[arjunrajlab]

The bug

A Cellpose-SAM run died with:

HTTP error 400: POST .../upenn_annotation/multiple
{"message": "data.coordinates must contain at least 1 items", "type": "validation"}

Annotations upload as one batch, so a single bad polygon took all 442 cells found in that frame down with it.

399df15 fixed the upload chokepoints for this — but only the code that runs after a shapely geometry exists. Construction itself was still bare, and that's where the cellpose-family workers build their geometry, before anything reaches a chokepoint:

• Polygon(contour) raises ValueError: A linearring requires at least 4 coordinates for a 1–2 point contour (a one-pixel-wide mask);
• a contour carrying a NaN/inf coordinate builds fine, then detonates inside .simplify() with GEOSException: Non-finite envelope bounds — which is not a ValueError, so it slips past a naive except ValueError.

Either one aborts the whole run, so one unusable mask out of hundreds still loses every good annotation in the frame. The failing run used Padding: -1 and Smoothing: 0.7, i.e. both transform paths were live.

The fix

New guarded helpers in annotation_utilities.annotation_tools, returning None/[] instead of raising:

Helper	Purpose
safe_polygon(coords)	Build a Polygon, or None. Rejects too-short contours and non-finite coordinates, and drops a third coordinate column (a 3D ring yields 3-tuples, which break the (x, y) unpacking used to build annotation coordinates).
safe_buffer / safe_simplify	None-tolerant, GEOS-error-tolerant padding and smoothing.
clean_polygon_coords(coords, padding, smoothing)	The whole pipeline: contour → annotation-ready rings, padding then smoothing (the order the cellpose family has always used).

geometry_to_polygon_coords also now rejects a NaN area (every NaN comparison is False, so area <= 0 let it through) and survives GEOS errors raised by is_valid/area.

Sweep

Every worker that builds a polygon from model output had the same latent crash:

• cellposesam, cellpose, condensatenet — run_model post-processing now goes through clean_polygon_coords. Behavior is unchanged for healthy outlines (same buffer-then-simplify order, preserve_topology=True as before).
• sam2_video, sam2_propagate, sam/sam2 automatic-mask-generator, sam/sam2 few-shot — guarded safe_* construction, preserving the skip-unusable behavior those functions already had for their no-contours case.
• sam2_automatic_mask_generator also indexed contours[0] without checking for an empty contour list — IndexError on a mask with nothing traceable.
• sam/sam2 few-shot computed training-area stats from user annotations, where one degenerate annotation aborted the run before inference started.
• worker_client — the upload chokepoint now uses safe_polygon, so it catches GEOS errors too (its except (ValueError, TypeError) did not); helpers are re-exported for workers.

Removed the now-unused shapely.geometry.Polygon imports (duplicated in several of these files).

Deliberately not covered: the cellpose_train / cellposesam_train region-polygon reads, where a degenerate user annotation can still abort a training run — a different surface (user input, not model output) and out of scope here. It's recorded in the hardening skill catalog, which gains the construction-time half of this failure mode.

Tests (TDD: red → green)

workers/annotations/cellposesam/tests/test_run_model.py is new — this path had no coverage. deeptile/cellpose are stubbed so it runs natively in the light venv, no GPU image needed. 6 of its 10 tests reproduced the crash before the fix; the other 4 pin existing post-processing behavior (negative padding erosion, MultiPolygon splitting, pass-through, smoothing actually simplifying).

annotation_utilities:  35 passed   (+17 new)
worker_client:         15 passed   (+3 new)
cellposesam:           23 passed   (+10 new)

Docker worker tests were not run — no Docker in this environment. The pre-existing few-shot worker test suites need the in-image annotation_client and fail to collect natively both before and after this change.

CELLPOSESAM.md gains a note that degenerate outlines are dropped and that a pinched object yields one annotation per piece, so a frame's annotation count need not equal its mask count.

---

Generated by Claude Code