[Sad Catapella B6][Scientific QA] Malformed explicit NGFF coordinate transforms are silently treated as unknown calibration #94

Scope and classification

Sad Catapella Batch 6 at d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed diagnostic gap with deliberate fallback behavior (there are tests that explicitly accept this). This issue proposes a stricter scientific QA policy, not a claim that pixel data are misread. Distinct from #73 (missing axes) and closed #39 (unknown-scale placeholder).

Evidence

• _ngff_scale(), lines 550–569 raises for malformed coordinate-transform arrays or scale dimension lengths, while ngff_physical_sizes(), lines 504–512 converts those invalid metadata values into empty PhysicalSizes().
• zarr_input._physical_sizes(), lines 325–335 catches every exception when reading calibration, logs only at debug, and returns unknown sizes.
• metadata._check_pixel_sizes(), lines 794–801 emits missing_pixel_size only if at least one image has a known comparable size. If all files contain malformed explicit calibration, there is no calibration-specific warning.
• Existing test_unusable_multiscale_level_scale_makes_sizes_unknown and test_transformations_that_are_not_a_list_give_no_sizes intentionally accept this behavior.

Counterexample / impact

An OME-Zarr experiment where each image explicitly declares coordinateTransformations as a dictionary instead of a list passes pixel preflight with physical_sizes=PhysicalSizes(). If every image has this malformed field, there is no missing_pixel_size or invalid_calibration warning. Scientifically valuable calibration metadata are silently discarded; the cause is absent from normal outputs, potentially making downstream analyses appear merely uncalibrated.

Proposed policy and acceptance

Differentiate (a) not provided size (validly unknown) from (b) explicitly provided but malformed transform (invalid metadata). Keep pixels readable if desired, but emit a path-specific, actionable invalid_physical_calibration issue in preflight and run, optionally block calibrated downstream outputs. Do not silently substitute 1 µm or require all images to be calibrated. Add mixed and all-malformed datasets, zero/negative scales, invalid arrays, unsupported transform types, valid unknown-scale marker, and checks that warnings survive manifest/report writing. Requires an explicit severity decision, no Numba.