Data-flow audit: minor findings (OME-Zarr scale, manifest flags, axis roles, CSV columns) #39

Data-flow audit 2026-10-29 (H1), MINOR. At v0.3.1.

• Unknown pixel size written as scale 1. In the OME-Zarr copies, physical_sizes.py:133-136 and zarr_export.py:414-425 write scale 1 with no unit. Leave the transform out, or record that the size is unknown.
• Manifest flags read with bool(). manifest.py:351,376,547 turn "false" into True. Accept JSON booleans only.
• No per-file axis roles in the manifest. Consumers rebuild them from the experiment's mapping, which is wrong for OME-Zarr files whose declared order differs. Record each file's roles.
• CSV columns ignored. The normalization v0.3.1 CSV reader ignores the plane_id, image_file and bit_depth columns and recomputes plane_id, so an inconsistent row is accepted. Check them.
• Labels above 2^63. uint64 masks with labels above 2^63 are accepted but cannot become an int64 label_id downstream. Refuse them.
• Version type in the contract. Contract §9 shows manifest_version/schema_version as numbers, but they are written as strings ("1", "4"). Say so in the contract.