[Sad Catapella N7][Scientific QC] Inconsistent ImageSummary channel dimensions can omit repeat checks or raise IndexError #56

Baseline and classification

Normalization Batch 7, f784c95e3049483f87466150ffc5764375fa631d. Confirmed unchecked public input dimensions and control-flow inconsistency, distinct from #23 (RGB type contract) and #44 (plane coordinates). No tests executed.

Evidence

• ImageSummary is a public dataclass with no __post_init__: the median, p_low, p_high, saturated_fraction, and exact tuples can have different lengths, and the declared dtype/fractions can be inappropriate.
• intensity_checks() lines 153–180 performs no consistency validation on independently supplied summaries.
• _repeat_shifts() lines 229–249 iterates range(summaries[0].channel_count) for all repeats and indexes s.median[channel]. If the first summary has one channel but later repeats have two, C1 repeat-shift QC is silently omitted. If the first has two and a later repeat has one, it may raise an unrelated IndexError.
• _image_outliers() lines 188–198 similarly uses each group's first summary's count, and _no_spread() lines 301–305 zips p_low/p_high with strict=True; inconsistent dimensions fail late or omit measurements.
• image_summaries() lines 165–171 validates channel counts for its own outputs, but that protection is bypassed for public direct ImageSummary inputs.

Fix and acceptance

At construction and/or the public QC entrypoint, require nonempty and consistent channel tuple lengths per summary, consistent channel cardinality among compared summaries (ideally across the supplied experiment), valid integer/float dtype and valid saturation fractions in [0,1] or None; ensure p_low <= p_high when both available, with explicit policy for absent values. Reject mismatches with source path, repeat and channel index rather than silently skipping or allowing IndexError. Test direct one-vs-two channel summaries in both orders, internally mismatched tuple lengths, empty/nonfinite channels, invalid fractions and ordinary image_summaries output. Numba offers no benefit.