[guijacquemet]

Fixes #47

Bug

At v0.3.3 (320f304), PlaneHistogram.__post_init__ filled a missing pixel_min with binning_min and a missing pixel_max with binning_max, and only then ran check_histogram(). A plane_histograms.csv row with valid pixels, a blank Intensity_MinIntensity_C0 and a populated Histogram_BinningMin_C0 was therefore accepted with a value made up from another column.

Fix

• histogram.py: the fallback is gone. If a plane has valid pixels, all four limits must be given. A plane with no valid pixels still has all four set to None.
• Two internal callers relied on the fallback. Both now pass the extrema explicitly:
  • _table_histogram.min_max, the min-max histogram for 8- and 16-bit planes;
  • _bin_pass with extrema is None, the min-max branch for floats and wider integers.
    No separate constructor was needed.
• integrity.py: the refusal now names each limit that is missing or not finite, for example pixel_min is missing or not finite.
• csv_io.py: if a channel has valid pixels and one of its limits is blank, the CsvFormatError names that column, for example Intensity_MinIntensity_C0 is blank, but the channel has 4 valid pixel(s).
• Docs: the contract's histogram invariants, README "Unreleased" notes, and an AGENTS.md map entry for tools/. The version is not bumped.

Tests

• In memory: each of the four limits set to None one at a time is refused, with the field named. A robust histogram with a missing pixel_min or pixel_max is refused. Computed min-max histograms carry their extrema.
• In the file: each of the four limit columns blanked one at a time is refused, with the column named. A channel with no valid pixels and blank limits is still read.
• Test helpers now pass their extrema: singleton, histogram(), and the v2 fixture reader, which reads the v2 pixel_min/pixel_max columns.
• On origin/main, the 10 new cases fail. The pixel_min/pixel_max cases fail because the value is silently filled in; the binning cases fail only on the new message.

Bitwise unchanged outputs

tools/dump_outputs.py is committed with this PR. It dumps the following for simulated planes of all supported dtypes, with float.hex for every float:

• histograms with the min-max, robust default and robust tight ranges (with tails), as whole planes, blocks and histogram_with_range_check (including the RangeCheck);
• plane_histograms.csv bytes (sha256) and the records read back;
• PooledHistogram statistics;
• image_summaries and intensity_checks;
• fit for 4 methods × 2 scopes.

The dump is 1788 lines. Both runs used the same environment: Python 3.13, the repo's dev deps, and the script from this branch.

Code	SHA-256
origin/main 320f304 (via PYTHONPATH=<worktree of main>/src)	d82a9f97dc39d7851f1cee0d22b85ac442f92acbaef7b02892ea55ba6a4e218b
this branch	d82a9f97dc39d7851f1cee0d22b85ac442f92acbaef7b02892ea55ba6a4e218b

The two dumps are also identical byte for byte (cmp).

To reproduce:

python tools/dump_outputs.py                                   # this branch
git worktree add /tmp/main origin/main
PYTHONPATH=/tmp/main/src python tools/dump_outputs.py          # main

Consumer impact (read-only git grep origin/main)

• morphobricks-preprocessing (79dd68e): tests/test_schema_v2_equivalence.py:134 builds PlaneHistogram(...) from v2 rows without pixel_min/pixel_max. It will be refused once preprocessing pins the next tag; it should pass pixel_min/pixel_max from the v2 columns, as this PR does in its own copy of that test. src/ only uses the type in annotations and calls compute_*, so it is unaffected.
• CellPatcher (40a68b1): only type annotations and compute_plane_histogram. Unaffected.
• MorphoFeature (1e6a610): no references.

Gate (local)

• ruff check .: all checks passed
• ruff format --check .: 43 files already formatted
• pytest: 622 passed, 2 skipped. The 2 skipped tests are tests/test_rgb.py, which need csbdeep; the CI csbdeep job runs them.

Note: histogram.py was already 864 lines before this PR, over the 600-line standard. This PR makes it slightly shorter but does not split it.

---

Generated by Claude Code