feat: navigator widget, label features and URL reading, with the phase 1 and 2 review fixes (napari plugin phase 3)

[aofei-liu]

Summary

Phase 3 of napari_plugin/PLAN.md, stacked on #6. Adds the sp-ops navigator dock widget (a lazily expanded tree of the store with checkboxes and an add button), label features from a table joined by a computed edge in sp-ops:relationships, contrast limits from the lowest pyramid level, and HTTP/URL reading with detection, child opening and parquet all going through the store or fsspec. The second commit fixes every finding from the delegated reviews of #5 and #6; they land here because the four PRs are meant to merge together today. Per-PR review-outcome comments on #5 and #6 list what was fixed and where.

Acceptance criteria

• a dock widget shows the store as a tree with checkboxes and an add button, so a user opens one well without re-dropping. test_navigator.py (tree model and an offscreen Qt smoke test)
• a merged collection with a computed edge from a labels element to a table on value/label gives the labels layer the table's obs columns as features. test_labels_layer_carries_table_features, test_merged_collection_opens_over_http
• contrast limits come from the lowest pyramid level. test_image_contrast_limits_come_from_the_lowest_level
• opening by HTTP URL is exercised against an example store served locally. test_url.py (synthetic store; the real stores are 1 GB and not served in tests)

Decisions

D19 points and shapes coordinates are taken as written and the offset goes to translate. D20 features only from computed key edges, obs read through zarr. D21 detection claims sp-ops stores or a dropped collection and stops at the .zarr component. D22 Qt-free tree model under a thin Qt widget. Recorded in PLAN.md.

Deviations from plan

• Rule 3: the widget was built by a parallel agent; its table test was pointed at the fixture's real table.
• Commit separation is by module rather than strictly by concern: images.py carries both the contrast-limits feature and two review fixes.
• tests/spops_store.py is 320 lines, over the soft limit, as the single fixture writer.

Verification

• cd napari_plugin && QT_QPA_PLATFORM=offscreen UV_NO_SYNC=1 .venv/bin/python -m pytest tests -q → 51 passed, 7 skipped. Baseline on the phase 2 branch was 38 passed, 6 skipped.
• With SP_OPS_PROCESSED_EXAMPLE and SP_OPS_RAW_EXAMPLE set → 58 passed.
• uvx ruff check --select F,B,UP,SIM src tests → clean.

Self-check: PASSED

📋 Phase 3 of napari_plugin/PLAN.md
🤖 Generated with Claude Code

[aofei-liu]

Follow-up after opening two collaborator stores written with ome-zarr-py (commits on this branch after 2650bf2).

Correction to how this was first described: the collaborator stores follow the RFC-8 text, and the two example stores this plugin was developed against are the transitional ones. RFC-8 defines three node types, collection, multiscale and singlescale, and says the Multiscale node's nodes "MUST be an array of Singlescale objects" and that this interface "replaces the multiscale metadata defined in the previous versions". docs/extension.md line 26 names the same three types. The open-questions page raised exactly this in Q18 and Q28. The example stores keep a 0.5 multiscales list inside RFC-8 collections, which the reader must still accept because napari-ome-zarr and every existing 0.5 image use it.

What the reader lacked: the RFC-8 form (no multiscales list, levels as singlescale nodes with per-level coordinateTransformations, axes only in the node's coordinateSystems), so it classified every image as unknown and every collection came back empty. Placement in a well-level scene was reported and ignored; points in the image pixel frame had no scale; one table uses AnnData's nullable-string encoding; one labels edge is declared on the plate.

Fixed: levels and axes are read from the RFC-8 form as well as the 0.5 list (D23); scale/translation/sequence scene transforms from any ancestor are padded onto the image axes by name (D17 revised); points sharing a collection with an image take its scale and translation (D19 revised); nullable obs columns decode; label features try every computed edge from every ancestor collection and skip endpoints that are not tables; table warnings fire only for a directly dropped table.

One note for the store writer: RFC-8 says a singlescale transform's input references the id of the singlescale node itself; these levels have no id and reference a per-level coordinate system instead. The reader does not use input, so nothing breaks, but a strict validator would flag it.

Result: biohub store root opens as 21 layers at the scene's 15600 µm offset with cell_seg features from the plate-level cells table; scallops store root opens as 14 layers with a 9-step round slider, scaled peaks and reads, and features on nuclei, cells and cytosol. Suite: 63 passed with all three store sets, 54 passed and 9 skipped without.