test: simulation-mode acquisition acceptance suite

[hongquanli]

Summary

End-to-end acceptance suite for simulated acquisitions — the safety net for the upcoming acquisition-engine refactor and protocol-v2 port. New software/tests/acceptance/ package (marker acceptance), driving MultiPointController in-process (no GUI, no QApplication — the controller uses plain threads and plain-callable callbacks) and asserting on observable artifacts only: files on disk and process state.

Scenario	Test	Pins
Full small acquisition	test_full_small_acquisition	2×2 FOV × 2 ch × 3 z ⇒ exactly 24 images in 0/, 12 rows in per-timepoint coordinates.csv, 4 rows in root CSV, acquisition parameters.json contents, .done markers, no ERROR lines in acquisition.log
Format parity	test_format_parity_ome_tiff_vs_zarr	same geometry under OME_TIFF and ZARR_V3 ⇒ 16 planes each, identical plane shape (4168×4168) and dtype (uint16)
Abort mid-acquisition	test_abort_mid_acquisition_then_reacquire	clean stop after abort; a second acquisition succeeds in the same process; zero JobRunner subprocesses survive close() (the #582 orphan surface)
Backpressure regression	test_zstack_completes_under_tight_byte_limit	12-image z-stack under a ~1.5-frame byte limit (~8× oversubscribed) completes instead of deadlocking; throttling verified engaged in logs
Multi-timepoint smoke	test_multi_timepoint_smoke	Nt=2 ⇒ 0/ and 1/ each with correct images, CSV rows, .done

Shared harness (harness.py + conftest.py) wraps the existing tests/control/test_stubs.py wiring; timeouts are deliberately generous for CI runners.

CI wiring

No workflow change needed: pytest's default collection picks up tests/acceptance/ under the existing python3 -m pytest step. Deselect locally with -m "not acceptance". Suite runtime: 5 passed in ~28 s locally (headless).

Behavior findings (pinned, not fixed — recorded per the session brief)

1. acquisition parameters.json is written at start_new_experiment() time, not at run_acquisition() time — callers that configure NZ/Nt after creating the experiment get stale values in the JSON. Tests configure before creating; noted in the harness docstring.
2. OME-TIFF output ignores the timepoint dir: files land at {experiment}/ome_tiff/{region}_{fov}.ome.tiff (one per FOV, series ZCYX), while individual-images mode writes under {experiment}/0/.
3. Zarr v3 output also lands at the experiment root ({experiment}/zarr/{region}/fov_{n}.ome.zarr/0, 5D TCZYX), not under the per-timepoint path that ZarrWriterInfo.get_output_path's docstring suggests.
4. The product's Zarr v3 stores are unreadable by the installed zarr-python 2.15 — tests read them back with tensorstore (the same library the product writes with), and the parity test importorskips it, matching test_zarr_writer.py.
5. FILE_SAVING_OPTION is star-imported by multi_point_worker, so runtime format switching requires patching both the _def attribute and the worker's frozen binding — the conftest helper does this; a future cleanup could move the worker to the control._def.X runtime-access pattern.

Test plan

• Full suite headless locally: python -m pytest tests/acceptance -q → 5 passed in 28.3 s
• Abort test 3× consecutive green; backpressure 2× consecutive green
• black --config software/pyproject.toml clean
• CI on this PR runs the suite under Xvfb via the existing pytest step

Session-B note: this PR is the Sync-2 gate for the parallel-session plan (Session C's code phase starts when this is merged).

🤖 Generated with Claude Code

[hongquanli]

The review-added premise guard caught a real config divergence on its first CI run — CI's pristine configuration_Squid+.ini crops the simulated camera to ~8.3 MiB frames vs ~33.1 MiB locally, so the hardcoded frame size overstated the byte pressure (12 frames were only ~2× the limit in CI, not the intended ~8×). Fixed in fd156084: the test now snaps one frame at start and derives the 1.5-frame byte limit from its measured nbytes, making the ~8× oversubscription hold under any machine config; the on-disk guard still verifies the premise after the run.

Note: that CI run also ended with the flaky post-green exit segfault that #595 root-causes (3 of master's last 15 runs fail the same way) and fixes via SQUID_PYTEST_HARD_EXIT — until #595 lands, any PR can flake at interpreter exit.