[guijacquemet]

Stacked on #190 (claude/discovery-header-additions, which adds the raw header channel names this PR compares). The base is that branch so the diff shows only this PR. Retarget to main after #190 merges. #188 is already merged into that branch.

What

Stage 1 design PR 8 (conformance P01-P03, P10, P12, P13, P23, and the comparison part of P26). Additive: no run, notebook or CLI command uses it yet, and nothing is written.

New module physical_check: check_physical(description, discovery, metadata, excluded=, roles=, mask_types=) returns a PhysicalReport.

• R_UNIT_NOT_FOUND (declared unit, no included image; P12) and R_UNIT_NOT_DECLARED. Units are matched by source folder labels, never by IDs.
• R_UNIT_NOT_HOMOGENEOUS (P01, P02): channel count, image type, interpreted axes or raw header name at a C index differ within a unit. The issue names the FOV, path and C index, and no per-FOV exception exists. It replaces R_CHANNEL_COUNT_MISMATCH and R_HEADER_CONTRADICTION for its unit.
• R_CHANNEL_COUNT_MISMATCH: a whole unit with another count than its declared channel list.
• R_HEADER_CONTRADICTION (P03, P26): raw name at Ci against the confirmed header_names of the channel at i. NFC of both sides, no trimming, no case folding, no similarity. Unnamed images need no evidence.
• R_IMAGE_TYPE_CONTRADICTS_HEADER (P13, P23): declared rgb or absent against the header's colour interpretation, in both directions. Pixels are never read, so P23(c) passes with no inference. OME-Zarr states no colour interpretation, so it cannot contradict and is recorded image_type_verified = False.
• R_MASK_TYPE_NOT_DECLARED (P10): only when the discovered mask types are passed. A declared mask with no files is valid. No mask representation is defaulted or read (Q5).
• layout_groups (Q3): units with an identical layout are grouped. Only a group whose channels are all header-named is batch_confirmable. Unnamed or partly named layouts need one confirmation per acquisition.
• Q7: homogeneity compares the axes with T left out, so time-lapses of different lengths, or one time point with no T axis, are one layout.
• Undecidable images are listed in unchecked, never passed: unreadable headers, and a channel count the header does not state (a plain stack) when no confirmed roles are given.

Choices not fixed by the Spec text (please check)

• An image type that contradicts the header also replaces the count and header-name checks of that unit. P23(a) and (b) expect only R_IMAGE_TYPE_CONTRADICTS_HEADER.
• Units declared rgb skip the header-name check, since R, G and B take no aliases (4.4b) and a name could never be confirmed.
• A unit whose images are partly named and partly unnamed is R_UNIT_NOT_HOMOGENEOUS (the literal "header names at one C index differ"). The Spec text on whether unnamed images count as differing is silent.
• Only T is left out of the axes comparison, as decided. Z is not: a unit mixing a stack with a single plane differs in axes and is reported. Decided by the owner (2026-11-07): a unit mixing Z-stacks with single planes stays "not homogeneous". No change needed.
• Not in this PR (other design rows): R_HETEROGENEOUS_LAYOUT_NOT_ENABLED and the mask-pixel check (PR 11), R_FOLDER_NOT_NFC and R_MASK_TYPE_NAME_INVALID (Discovery and header additions for the experiment description (Stage 1, PR 6) #190), N35 reopening (PR 9), and the plain-language texts.

Tests

tests/test_physical_check.py, 23 tests, built on the vendored fixtures v01, v07, v11, v12, v28 and v31. They write real TIFFs and change one image or declaration. On the base branch (#190) the file fails at collection (ModuleNotFoundError: morphobricks_preprocessing.physical_check). Here: 23 passed.

python tools/check.py: 3863 passed, 4 skipped, code rules clean. The 4 skips are environment ones (case-insensitive file system and file-mode tests, unrelated to this change).

---

Generated by Claude Code