[guijacquemet]

Summary

Part of #39. Five of its six checklist items are fixed here; the CSV-column item is left for the normalization brick (see below).

• Unknown pixel size written as scale 1 (zarr_export.py, physical_sizes.py). OME-NGFF 0.4 requires a scale on every axis, so the transform stays and 1 is still written. What changes:
  • The multiscales entry's metadata now lists those axes as "unknown_scale": ["z", "y", "x"].
  • ngff_physical_sizes returns None for them, so 1 is never read as a size.
  • PhysicalSizes.known(role) is added.
  • copy_units_current also compares this list, so copies written before it are converted again. Their rows do not change.
• Manifest flags read with bool() (manifest.py). missing_as_singleton, convert_tiffs, read_in_blocks and channels_from_rgb_samples must now be JSON booleans. Any other value ("false", 0) raises ManifestFormatError, and the run then rebuilds as for any unreadable manifest. Before, "false" was read as True.
• Per-file axis roles (manifest.py). FileEntry.roles and the file entry key roles hold one role or null per axis, from the file's own mapping. This is correct for OME-Zarr images whose declared order differs from axis_mapping. It is an optional key, so the manifest stays version 1; it is null in older manifests and filled on the next run, which writes fresh entries.
• uint64 labels above 2^63 (masks.py, cli.py). read_mask raises ValueError for such labels. The masks command prints the message and exits 1.
• Version types in the contract. Contract §9 now says manifest_version and schema_version are JSON strings ("1", "4").
• Docs: contract §2, §9 (new Flags row, roles), §10 (Scale row) and §15; the AGENTS.md manifest summary.

Not done: CSV columns ignored

The plane_id, image_file and bit_depth columns are ignored and plane_id recomputed by morphobricks-normalization's read_plane_histograms (v0.3.1). That reader lives in that repository, which I was not to modify here. It should be fixed there, and this brick can then pin the new tag. Within this brick, reused rows are already guarded by the manifest's csv_sha256. So this PR does not close #39.

Notes for review

• unknown_scale is a key of this brick's own, inside the NGFF metadata object, which is free-form. Other OME-Zarr readers will still see scale 1 unless they check it. The only spec-valid alternative is to keep writing 1 without saying so.
• The negative-label check (A signed mask with a negative label makes the background an object #35, PR Refuse negative mask labels (background was counted as an object) #40) and the cli ValueError handling also touch read_mask and _masks. Whichever PR merges second needs a small rebase. Both checks belong in the same place.

Tests

New tests:

• tests/test_manifest.py::test_flags_must_be_json_booleans (16 cases)
• tests/test_manifest.py::test_each_files_axis_roles_round_trip_and_are_none_when_absent
• tests/test_manifest.py::test_file_entry_for_records_the_files_own_roles
• tests/test_zarr_input.py::test_the_manifest_records_each_files_own_axis_roles (TIFF ZCYX next to OME-Zarr czyx)
• tests/test_zarr_export.py::test_unknown_pixel_sizes_are_marked_not_read_as_1
• tests/test_zarr_export.py::test_known_pixel_sizes_are_not_marked
• tests/test_masks.py::test_uint64_labels_above_the_int64_range_are_refused

Results:

• All 22 new test cases fail on main.
• Full suite (Python 3.13, zarr 3.4.2, notebook extra installed): 1073 passed, 1 skipped. The skip is test_discovery.py:472: running as root ignores file modes.
• ruff check . and ruff format --check . are clean.
• Not run: zarr 2.18 / Python 3.11, Windows, macOS.

---

Generated by Claude Code

[guijacquemet]

Review follow-up, pushed as 44b101b:

• Strict writer: the manifest writer is now as strict about flags as the reader. missing_as_singleton, convert_tiffs, read_in_blocks and channels_from_rgb_samples are checked with the same _flag rule when the manifest is written (null is still allowed for channels_from_rgb_samples). A value that is not a bool ("false", 0, 1, numpy.bool_) raises a ValueError naming the key, before any file is written. Nothing is coerced. Contract §9 is updated.
• Tests:
  • test_the_writer_refuses_flags_that_are_not_booleans: 16 cases, all fail on 699e5b4.
  • test_the_writer_keeps_null_for_an_unknown_rgb_flag.
• Checks: pytest 1090 passed, 1 skipped (test_discovery.py:472, running as root). ruff check and ruff format --check are clean.
• Merge note: this PR conflicts with Make differing channel names at one channel number an error #42 (docs/contract.md, manifest.py), Give masks a plane identity (Z/T); refuse 2D masks of multi-plane images #43 (docs/contract.md, masks.py, tests/test_masks.py), Refuse negative mask labels (background was counted as an object) #40 (cli.py, masks.py, tests/test_masks.py) and Report a malformed OME-Zarr labels group as unreadable_mask #41 (docs/contract.md, tests/test_masks.py). The ones in masks.py and cli.py were already there before this follow-up. Once Make differing channel names at one channel number an error #42 lands, its new channel_names_trimmed flag should go through _flag too.

---

Generated by Claude Code

[guijacquemet]

I added 9128250, so run_experiment refuses a non-bool missing_as_singleton/convert_tiffs before reading anything. Before this, the stricter writer only failed after the full run; a regression test covers it. Then I merged main (#40, #42, #43) in:

• manifest FileEntry keeps both channel_names_trimmed and roles;
• read_mask keeps Give masks a plane identity (Z/T); refuse 2D masks of multi-plane images #43's plane read, Refuse negative mask labels (background was counted as an object) #40's negative-label check and Data-flow audit minor findings: unknown scale, manifest flags, per-file roles, uint64 labels #44's uint64 limit;
• contract §2/§9/§10/§15 rows were combined (unknown_scale with the unlabelled-channel copy rule, and the uint64 note in the Z/T header checks);
• all tests were kept.

1120 passed, 1 skipped (test_discovery.py:472, root), ruff clean, CI 7/7 green on 3e833c8. Merging.

---

Generated by Claude Code