[Sad Catapella B5][Scientific integrity] Source files may change between fingerprint capture and histogram publication #90

Baseline / classification

Adversarial audit Batch 5 at d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed TOCTOU validation gap; no concurrent-acquisition reproducer run.

Evidence

• file_entry_for() lines 415–445 fingerprints source images while constructing the entries before image pixel reads.
• _read_file() then reads/accumulates pixel data, potentially over several passes and many files. run_experiment() lines 636–690 publishes the earlier fingerprint without validating the source again. This can be problematic for active time-lapse acquisitions or writable OME-Zarr datasets.
• check_sources() detects later metadata changes when explicitly called but cannot establish snapshot consistency during the original run.

Deterministic test

Instrument iter_xyc_planes to mutate one source TIFF's valid pixel data, keeping geometry/dtype, after the initial fingerprint or between histogram passes. Verify that a successful run never records a fingerprint of one image generation while its table describes another generation; assert explicit refusal and no publication. Include Zarr chunks, source file deletion, and remote-writer cases. For a simple TIFF, an actual mtime increase may be caught on next run; that does not protect the published run from an inconsistent snapshot.

Fix

Capture a beginning-and-end signature for every input whose pixels were read; before committing compare it to the pre-read signature and block/retry if changed. For stronger correctness across a multi-pass read, require immutable input snapshots / stable versions or per-chunk content hashing where practical (#86). Document that directory timestamps cannot detect all mutations. Avoid silent replacement of source fingerprints after work. Benchmark signature validation overhead. No Numba value.