[guijacquemet]

Owner decision on #37: strict. Parameters.from_columns now refuses a row whose normalization_scope differs from settings.scope, just as it already refuses a different normalization_method.

Fixes #37

Changes

• parameter_checks.check_scope(scope, expected, scopes) refuses two kinds of scope:
  • an unknown scope, with the same message as in Bucket A (2/2): refuse invalid parameters and channel indices; report unchecked channels #62;
  • a scope other than the settings'. That message names both scopes and asks to read the parameters with the settings of the run that wrote them.
• fitting._scope is removed and from_columns calls check_scope. This keeps fitting.py at 591 lines; adding the check inline would have taken it to 608.
• Tests:
  • test_columns_round_trip_exactly reads its patch rows with NormalizationSettings(scope="patch").
  • New: every row × settings scope pair (3 accepted, 6 refused).
  • New: per-time-point and group values read back unchanged.
• Docs: contract §8 from_columns row and §9.1, AGENTS map.

Evidence

• ruff check . and ruff format --check . pass.
• pytest: 610 passed, 2 skipped (csbdeep, which is not installed locally).
• Mutation check: with the new comparison disabled, 6 tests fail.
• No other caller in this repo reads columns with settings of another scope.

Consumer impact (CellPatcher, when it repins past v0.3.2)

CellPatcher currently pins v0.3.2, so nothing changes for it until it repins.

cellpatcher/normalization.py::parameters_from_row (origin/main 79adc3a, lines 270-276) builds NormalizationSettings without scope. Against this branch, two CellPatcher tests fail with normalization_scope is 'patch', but the settings use scope 'image':

• test_read_patch_table.py::test_an_exported_run_reads_back_as_its_patches
• test_rgb_normalization.py::test_rgb_raw_export_and_lazy_delivery_give_the_eager_arrays

CellPatcher must pass the run's scope:

brick = mbn.NormalizationSettings(
    image_type=settings.get("image_type", "fluorescence"),
    method=settings.get("method", "percentile"),
    percentile=tuple(settings.get("percentile", (1.0, 99.8))),
    scope=settings.get("scope", "image"),
)

I tried this in a scratch copy, not in the CellPatcher repository. Both failures are fixed, but test_normalization.py::test_raw_patches_of_a_removed_scope_still_normalize_on_load then fails. That test's run_config.json holds the removed scope condition, so NormalizationSettings(scope="condition") raises.

Real rows from a condition- or technical_repeat-scope run hold that scope in normalization_scope. Since #62, from_columns refuses them as an unknown scope whatever settings are passed. For those runs, CellPatcher has two options:

• refuse with its removed-scope message;
• build Parameters from the pair columns itself, as it already does for the legacy ;-separated rows. Parameters(...) does not check the scope.

---

Generated by Claude Code