[Sad Catapella B6][Scientific calibration] Reused OME-Zarr copies are not checked for numeric pixel-scale correctness #92

Scope and classification

Sad Catapella Batch 6; pinned main commit d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed source-path omission, with a concrete metadata-only counterexample; end-to-end tests were not run. Distinct from #87 (integrity of converted pixel chunks) and #88 (channel labels).

Evidence

• copy_units_current(), lines 538–545 reads multiscales[0].axes and compares only units and the proprietary unknown_scale marker. It never verifies datasets[0].coordinateTransformations (nor any higher-level scales).
• run_experiment(), lines 550–559 uses this unit-only predicate to reuse a previously converted TIFF's OME-Zarr copy without reconversion.
• OmeZarrWriter.attributes(), lines 410–445 writes the physical scale for each pyramid dataset. Existing unit-freshness regression changes axis units, but no test changes numeric scales while leaving units intact.

Counterexample

1. Convert TIFF with X=Y=0.25 µm/pixel to an OME-Zarr copy; save successful run/manifest.
2. Edit .zattrs multiscales[0].datasets[0].coordinateTransformations[0].scale X from 0.25 to 0.5, retaining axes[*].unit, unknown_scale, source TIFF and channel labels.
3. On an unchanged rerun with convert_tiffs=True, copy_units_current returns True, so the copy is reused with a false 0.5 µm/pixel scale.

Impact

Downstream physical distances, areas and speeds from the converted image can be wrong even though pixel data and histogram counts are correct. Particularly important because MorphoBricks downstream bricks use calibration for spatial features.

Fix / acceptance criteria

Have the copy validation compare the effective numeric scale on level 0 for every known physical axis to the source PhysicalSizes with an explicit tolerance policy, and validate expected scale factor 2^level on X/Y in all pyramid levels. Respect/verify unknown_scale markers; reject missing/nonfinite/misordered/incorrect transformations instead of guessing. Consider a cryptographic metadata fingerprint/generation marker as part of #87, but do not rely on unit equality. Add altered level-0 and level-1 scale fixtures, unknown units, valid times, mixed X/Y units and no-change control. No Numba role; metadata validation is low-cost.