[guijacquemet]

Bucket-A integrity fixes, parameter side. The statistics side is #61. The two PRs merge cleanly; combined, fitting.py is 599 lines.

Fixes #36, fixes #38. Refs #37, refs #57. Both are partly fixed and stay open; the reasons are below.

Changes

• New parameter_checks.py ([Sad Catapella N3][Correctness] Direct Parameters/NormalizationParameters accept invalid offsets and scales #36): Parameters and NormalizationParameters refuse, when constructed:

  • a non-finite offset;
  • a scale that is not finite and above 0.

  The message names the channel or method. Non-finite pixels still pass through apply unchanged. fit still gives its existing zero-spread message.

• Channel indices ([Sad Catapella N3][Data integrity] Parameter channel indices silently truncate fractions and booleans #38): Parameters.channels and select accept Python or NumPy integers ≥ 0 only. Bool, 1.9, 1.0, "1", -1 and None are refused instead of being truncated by int().

• Scope column ([Sad Catapella N3][Scientific integrity] Parameter deserialization does not verify normalization_scope against requested settings #37, partial): from_columns refuses a normalization_scope outside biological_repeat/image/patch, for example experiment or a blank value.

  • It does not compare the row's scope with settings.scope. CellPatcher's parameters_from_row builds settings without a scope (so the default is image) and reads rows of every scope. The repo's own test_columns_round_trip_exactly also reads patch rows with default settings.
  • A strict default would therefore break CellPatcher's export and delivery for biological-repeat and patch runs. That needs an owner decision, plus a one-line CellPatcher change that passes scope=.

• Unmeasured channels ([Sad Catapella N7][Scientific QC] Entirely nonfinite channels produce no warning and are labelled Histogram_Exact=true #57, partial):

  • ImageSummary.exact, and so Histogram_Exact_C<i>, is now None instead of True for a channel without finite pixels.
  • New unchecked_channels(summaries) returns one channel_without_finite_pixels warning per affected image and channel.
  • It is deliberately not added to intensity_checks. Preprocessing's notebook does counts[issue.code] += 1 over a fixed four-code dict, so a fifth code would raise KeyError there. Preprocessing has to adopt the new function to present the warning.

• Docs: contract §7.1, §8 (Validity row, scope), §9.1, §9.2; AGENTS map; README.

Evidence

• Regression tests: tests/test_parameter_checks.py, 31 tests. Against main, 24 fail; the other 7 check that valid inputs are still accepted.
• Bitwise comparison: the same 5,769-output dump as Bucket A (1/2): refuse corrupt statistics: blocks, counts, ranges, tails, plane identity #61 gives the same SHA-256 f7604ef1…c091417 on main, this branch and the combined merge. The only intended output change is exact for channels with no finite pixels, which are not among the valid inputs in the dump.
• Gate:
  • ruff check . and ruff format --check . pass;
  • --ignore-noqa reports the same 3 pre-existing E402 as main;
  • pytest: 505 passed, 2 skipped (csbdeep), 0 failed;
  • main + Bucket A (1/2): refuse corrupt statistics: blocks, counts, ranges, tails, plane identity #61 + Bucket A (2/2): refuse invalid parameters and channel indices; report unchecked channels #62: 600 passed, 2 skipped.
• Consumers: CellPatcher, MorphoFeature and preprocessing have the same pass/fail sets as against main (see Bucket A (1/2): refuse corrupt statistics: blocks, counts, ranges, tails, plane identity #61). Parameters, select and from_columns are used by CellPatcher (normalization.py:276,284,372,437,609) with valid integer channels and positive scales, so nothing breaks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G3qp9AxjpGPw9xfWAUJNqc