[guijacquemet]

Summary

The owner's decisions of 3 November 2026, and version 0.4.0rc5.

Masks: a tie between parents gives no parent (plan §12 Q74)

• mask_scan.pair_counts: before, np.lexsort((-shared, kid)) was stable, so when two or more parents shared a child's largest overlap exactly, the child went to the parent with the lowest label index. That was the only tie path: the same_objects counts come from a maximum matching, whose size does not depend on label numbering (Q67), and the hierarchy is sorted by name.
• New pure module mask_pairs.py with best_parents (gives each child's largest overlap and whether it is tied), plus one_to_one_matches and match_counts, moved from mask_scan so that it stays under 600 lines. mask_scan still re-exports one_to_one_matches. The layer is declared pure.
• A tied child has no parent and is never inside. If the tied share would make it inside (at least inside_pixel_share), it is counted in the new evidence ambiguous_parent. A tie below that share is split_between_parents, as before (it gets no parent either way). Every child is counted in exactly one of inside, ambiguous_parent, split_between_parents or outside_parents.
• MASK_REPORT_VERSION 2 -> 3.
• mask_text: new mask_ambiguous_lines. mask_report_summary (the masks command) has an "Objects given no parent" section that covers every pair, including pairs whose relation is none. The relationship table has a new last column. The notebook's Describe the masks card shows the lines.

TIFF calibration (physical_sizes, tiff_size_values)

• An ImageJ TIFF with a unit but neither XResolution nor YResolution now warns invalid_physical_calibration with the reason "X, Y: the unit 'um' is given but neither the XResolution nor the YResolution tag is (a unit without a resolution)". Before, it read as uncalibrated with no reason.
• A TIFF with no pages (tifffile opens it and only logs) gives "the TIFF file has no pages (no image)". Before, it read as uncalibrated with no reason.
• A file with neither a unit nor resolution tags is still uncalibrated.

Docs

• Contract §15.1 (report version 3, the tie rule, the ambiguous_parent row) and §11 (the two new causes of invalid_physical_calibration).
• User guide (mask step, masks command, message table) and AGENTS.md repo map.
• Plan: Q66 confirmed by the owner. Q73 decided: neither a stated reshape nor a reorder in v0.4, post-matrix issue convert: accept a stated page reshape (e.g. --pages CZ=2x3) and/or reorder copies into the common order (plan Q73) #144. New Q74 for the tie-break. PP-13 follow-up (a) marked done. rc5 added to the release checklist.
• CHANGELOG: a new 0.4.0rc5 section covering Fix cross-brick test findings F1, F8, F9, F10, F11 #139, Fix the rc4 external review: mount points, hard links, commit fsync, content verification #140, Round-2 audit fixes: strict v2 position_id (A-m5), physical-size read errors (A-m6), exact unit conversion (B-m7) #142 and v0.4.0rc5: tie-break gives no parent, ImageJ unit without resolution, plan Q66/Q73/Q74, changelog #143, with migration notes (mounts refused, position_id required in v2, µm conversion within 1 ULP, --verify-content, the tie-break), plus entries under 0.4.0. Migration guide: §5, §10, §11 and §13.

Version: 0.4.0rc5 in pyproject.toml, the notebook (current_version, RELEASE_TAG and the Step 0 callout), notebook_latest_versions.yaml, the README and the user guide. Historical rc4 mentions are kept: the rc4 manifest fixture and its test, the contract's "rc4 keys", and the plan's rc4 entries.

Consumers of the tie-break fields (read-only grep, 3 November 2026)

• CellPatcher: checked at HEAD 438e6d4 (branch claude/fix-audit2) and at origin/main 79adc3a.
• MorphoFeature: checked at HEAD 9e66ca6 (branch claude/request-check-hook) and at origin/main 142c448.
• Searched for: inside, parents_with_*, split_between_parents, outside_parents, inside_parent_larger, ambiguous_parent, mask_relationships, pair_counts, scan_masks, MaskReport.
• Result: no code reads any of them.
  • The only hits are a pytest parameter name inside in CellPatcher tests/test_notebook_session.py.
  • In MorphoFeature, prose in docs/contracts/provider-api.md uses "inside" for region masks, and D79 in the decision ledger says "the mask scan lives in preprocessing".

Decisions

• Owner, 3 November 2026: the tie-break, the ImageJ unit without resolution, Q66 confirmed, and Q73's follow-up.
• Within those decisions, my choices: a tie below inside_pixel_share stays split_between_parents, not ambiguous_parent, because it gets no parent under either rule. The no-pages case is reported as invalid rather than raised.

Notes for review

• child_pixels_with_parent_label compares label values, as designed (it is the grouped_by evidence). So the permutation test applies one label map to both masks; independent maps would change that count legitimately.
• The new tests were checked against the old behaviour: with best_parents patched to report no ties (the old lowest-index choice), 7 of the 9 tie tests fail. Against the old source, the 3 new TIFF tests fail.
• ruff check --ignore-noqa gives 60 findings, the same 60 as origin/main, all existing noqa'd entries. tools/code_rules.py and its baseline are unchanged (SHA cc04a420).
• Not verified: the notebook in a real JupyterLab or Colab (tested headless only).

Tests

Local, Python 3.13, .[dev,notebook,zarr] (zarr 3): 2197 passed, 4 skipped, 0 failed.

• The 4 skips are platform skips: 3 for "this file system tells names apart by case", 1 for "permissions that ignore file modes".
• ruff check ., ruff format --check . and python tools/check.py --rules all pass.
• CI: see the checks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3qp9AxjpGPw9xfWAUJNqc

[guijacquemet]

Independent adversarial review of the rc5 PR, starting at head 83ce987. Verdict: ready to merge after one documentation fix (e5e6a74).

Tie-break (Q74). It holds up.

• Each child is counted exactly once, in one of inside, ambiguous_parent, split or outside. The tests check this on random masks.
• Ties are found by comparing integer pixel counts (shared == most), so the test is exact. A 50 vs 49 near-tie stays inside.
• I relabelled the parent and child masks independently, with inside shares of 0.5, 0.34, 0.25 and 0.2, over 2,000 random cases containing 1,306 ambiguous children. Every count was unchanged except child_pixels_with_parent_label, which is defined by comparing label numbers. That test is now in the suite (test_the_counts_are_the_same_when_each_mask_is_numbered_on_its_own).
• A three-way tie is split at the default share and ambiguous_parent at 1/3.
• The same_objects matching code moved verbatim to mask_pairs, and the re-export in mask_scan is a single import.
• A tie below inside_pixel_share staying split_between_parents is consistent with the owner's rule. Such a child already had no parent, because no parent is assigned below the share, and its counts never depended on the numbering. ambiguous_parent means "would be inside, but which parent is undetermined". Moving those children would change split for cases the decision did not cover. At the default of 0.5, ambiguous means exactly half in each of two parents.
• Mutation testing: 17 mutants, 16 killed. The survivor (first versus last tied host) is equivalent, because a tied host is never used.

ImageJ / no pages.

• Both cases are warnings. A file with neither a unit nor a resolution stays uncalibrated with no reason.
• check, run and RunInputs agree for the unit-without-resolution case. I checked this end to end: one warning, and the image is recorded as uncalibrated.
• Fixed (documentation, important for the release): in check and run, a TIFF without pages is the error unreadable_tiff, as before. Its header has no series, so the new reason is reached only through read_physical_sizes. The CHANGELOG, contract, migration guide and user-guide table said it was invalid_physical_calibration. They now say where each one is reported, and a test pins the folder-check behaviour. The CHANGELOG and user guide also said every tie is ambiguous_parent; they now state the share condition.

Release.

• 0.4.0rc5 is consistent across pyproject, the notebook (current_version, RELEASE_TAG, the Colab text), notebook_latest_versions.yaml, the README and the user guide.
• The notebook follows the rules: # @title on every code cell, the Colab guard, and no magics.
• The CHANGELOG covers Fix cross-brick test findings F1, F8, F9, F10, F11 #139, Fix the rc4 external review: mount points, hard links, commit fsync, content verification #140, Round-2 audit fixes: strict v2 position_id (A-m5), physical-size read errors (A-m6), exact unit conversion (B-m7) #142 and v0.4.0rc5: tie-break gives no parent, ImageJ unit without resolution, plan Q66/Q73/Q74, changelog #143, matching git log v0.4.0rc4..HEAD --merges.
• The wheel builds as 0.4.0rc5, contains mask_pairs.py, and imports cleanly in a fresh venv.

§9.

• The touched files are within the limits (at most 570 lines).
• check.py --rules passes, and the code rules baseline file is unchanged. tools/code_rules.py is unchanged (cc04a420).
• ruff check --ignore-noqa is clean on every touched source file.

Gate on e5e6a74:

• ruff check and ruff format: clean.
• pytest with .[dev,notebook,zarr]: 2201 passed, 4 skipped (three because this file system is case-sensitive, one because the tests run as root), 0 failed.
• CI: 7/7.

---

Generated by Claude Code