B2: fix the remaining test_main failures

[trissim]

Fixes the tests/integration/test_main.py failures left after B1 (#1181). Depends on OpenHCSDev/PolyStore#34; the submodule points at that branch.

Causes and fixes

0. Test isolation (cross-talk). test_main hard-coded its ports: 7777 for the execution server, 5555/5556/5557/5559 for viewers, and Fiji's default 5565. On Linux the transport is IPC, so these ports name sockets in ~/.openhcs/ipc, which every checkout shares. Concurrent runs talked to each other's servers and viewers. During this work a viewer from another worktree rejected this checkout's metadata (Undeclared viewer metadata fields: ['source_channel_axis']).

• A new free_port_pair fixture in tests/integration/conftest.py allocates pairs through DataControlPortPairAuthority, the transport owner that binds them. It starts the search from an OS-assigned port, not the default port every run would pick first. At teardown it shuts down whatever still listens on those pairs.
• A new autouse reap_spawned_processes fixture marks each test's environment. Children inherit the marker, so at teardown it can kill the process group of every marked process still alive, pass or fail. This includes setsid-detached viewer and execution servers, which ViewerStateManager never stopped. The coordinator reported 38 leaked Fiji servers from an earlier run; after this fixture, ps -eo ppid,args | awk '$1==1 && /os.setsid/' is empty after every run.
• All parametrizations of one microscope used to share one tests_data plate directory. A zarr run converts its input plate in place (adding a second main subdirectory) and leaves outputs behind, so the next parametrization failed with Multiple OpenHCS metadata subdirectories are marked main. Each test now gets a fresh plate under its own tmp_path.
• Deleted the disabled _test_main_with_code_serialization and the ignored sequential_config parameter of create_test_pipeline.

1. zarr-OpenHCS: Zarr conversion output has no declared source projection. Broken by ccfef5f. update_metadata_for_zarr_conversion looked up store-relative paths (A01_….tif), but source projections are keyed by plate-relative virtual paths (TimePoint_1/A01_….tif). It now prefixes the original subdirectory. The existing unit test used "." as the subdirectory, where the two paths are the same. New regression test: tests/unit/test_zarr_input_conversion_subdirectory.py.

2. Other zarr cases: the results entry was missing required fields. Broken by 1f15250, on top of #134 (8224363). Step-time publication admits a result directory when exists() sees its saved files. The zarr backend answers that through disk passthrough. Completion writes the header fields only when is_dir()/list_files() find outputs, and zarr did not list passthrough files. The result was a results entry with projection fields only. Fixed in PolyStore: zarr list_files/is_dir now report passthrough saves. PolyStore#34 also fixes zarr delete(), which kept its own .json/.csv/.txt list, so replacing a saved .roi.zip on a re-run failed. New regression tests are in PolyStore and in tests/unit/test_persisted_roi_only_publication.py.

3. napari cases: ack timeouts and Saved images lack typed produced addresses. Broken by 7bc7816. It called Layer._make_slice_input, which the installed napari 0.9.2 does not have; slicing moved to Layer._slicing_state. Every scheduled layer update raised inside the viewer, so it never acked and producers timed out after 30 s. The "lack typed produced addresses" error is a knock-on effect: the failed steps left saved images without produced records. Fixed with a one-line change in napari_viewer_server.py. Regression test against a real ViewerModel: tests/unit/test_napari_native_frame_applied.py.

Coordination

• G6 (G6: domain-blind viewers #1178): one line in openhcs/runtime/napari_viewer_server.py (native_frame_applied). The G6 branch still has _make_slice_input. Separately, test_napari_reset_transitions.py fails 26 cases on _indices_view, another napari API that 0.9.2 removed. Those tests hang on main's viewer server. I did not fix them; they belong to G6.
• G3 (G3: tensor payload with declared axes #1180): no payload modules touched.
• Out of scope: detached viewers rewrite QT_QPA_PLATFORM=offscreen to xcb by design (QtPlatformName.OFFSCREEN is non-interactive). Running viewer modes with a display therefore opens real windows.

Results (serial, nice -n 19, worktree --basetemp, QT_QPA_PLATFORM=offscreen, DISPLAY/WAYLAND_DISPLAY unset)

• Non-viewer modes, 12/12 pass: disk and zarr × ImageXpress, OperaPhenix and OpenHCS × direct and zmq.
• Viewer modes, 36/36 fail from environment limits. These are the napari, fiji and napari+fiji modes. With no display, napari's viewer process dies at startup (no Qt platform plugin can start). Fiji runs headless and fails settlement with an InvocationTargetException. This machine has no Xvfb.
• Earlier runs with the display set, before the coordinator asked for headless runs: these passed after the fixes: disk-ImageXpress direct-napari, disk-ImageXpress zmq-none and zmq-napari, disk-OpenHCS zmq-napari, zarr-OpenHCS direct-napari, and zarr-ImageXpress direct-fiji.
• Official30 parity (CI command and env): 29/30. The failure is the known baseline cp_tutorial_3d_monolayer (.pkl serialization).
• Unit tests: PolyStore suite 594 passed. The touched OpenHCS unit files pass except the 26 G6 napari failures above.

B2 · PR #1182 · guards n/a · production −6 +9 (PolyStore −3 +32) · tests −219 +337 (PolyStore +38)

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.