[Sad Catapella N6][Scientific data integrity] RGB group fitting silently pools FOVs with incompatible integer pixel dtypes #51

Baseline and classification

Normalization Batch 6 at f784c95e3049483f87466150ffc5764375fa631d. Confirmed missing cross-record dtype constraint; whether a library-level guard is required is a scientific API ownership decision because preprocessing currently enforces experiment-wide dtype consistency. No package tests executed.

Evidence

• PlaneRecord._check_channels() enforces dtype equality within each XYC plane, not across records in a fitted group.
• fitting.fit() lines 357–395 accepts records with the same channel count, calls check_rgb(channel_count, _dtypes(items)) then pools all R/G/B intensity distributions.
• _dtypes lines 445–448 returns the set of distinct dtypes, but check_rgb() lines 77–100 only asserts every dtype is an integer: it does not require one common dtype across FOVs. Consequently uint8 and uint16 are both accepted together. This is distinct from [Sad Catapella B7][Type validation] RGB check accepts int32/int64 inputs despite stating 8- or 16-bit only #23, which addresses unsupported individual types (e.g. uint64).
• The normalization contract assigns experiment-wide dtype validation to preprocessing, but the public fit API is independently callable. RGB tests sample one dtype per test, not mixed types.

Deterministic fixture and impact

Create two valid PlaneRecord objects for different FOVs, each three-channel: one RGB uint8 with channels around 100–200, another RGB uint16 with channel values around 20,000–50,000; call fit([row8,row16], NormalizationSettings(image_type='rgb', scope='biological_repeat')). All samples are pooled into one shared reference as though they had comparable units/dynamic ranges. The result's quantiles and applied normalization can be dominated by the 16-bit FOV; no dtype inconsistency is reported.

Fix / acceptance

Decide whether the normalization library should enforce one dtype for every plane in an RGB fit (recommended), or require a prevalidated typed group descriptor and explicit caller proof. Never silently convert mixed dynamic ranges; do not infer/scalewise reconcile bit depth. Report conflicting source FOVs and dtypes before pooling. Regress single uint8, single uint16, mixed 8/16, mixed signed/unsigned 16, diverse yet equally typed image sizes and multi-timepoint grouping. Also assess whether this check should apply to fluorescence fit as a separate policy, keeping ownership with the relevant experimental contract. Benchmark error preflight overhead, negligible compared with histogram pooling. No Numba role.