[Sad Catapella N6][Performance/Validation] RGB fit computes full image histograms before rejecting invalid channel count or pixel dtype #52

Baseline and classification

Normalization Batch 6 at f784c95e3049483f87466150ffc5764375fa631d. Confirmed avoidable workload and late input rejection, runtime cost not measured.

Evidence

• fit() lines 357–379 calls _pixel_histograms(statistics, channel_axis) unconditionally for pixel-array inputs and only afterward calls check_rgb(channel_count, _dtypes(items)).
• _pixel_histograms() lines 436–442 constructs a complete histogram for each supplied channel, using np.take copies. For a four-channel uint16 FOV the calculation may include large per-channel arrays and multiple passes before a simple channel-count check rejects the RGB input. For invalid float32 RGB data, float histogram/range-check processing is similarly wasted.
• RGB tests assert the errors but never assert they happen before histogram extraction. Prior performance issue #40 concerns unnecessary copies for valid pixel-array fits; this issue concerns needless full-image computation on invalid RGB input.

Fix / acceptance

When settings.image_type == 'rgb' and statistics is a raw array, check validated channel-axis cardinality and raw dtype before _pixel_histograms. For record inputs, validate first record/channel_count and group dtype consistency before constructing pooled histograms. Ensure invalid channel_axis errors remain actionable, empty planes still report NoValidPixelsError as before, and the preflight policy agrees with #23's resolved RGB dtype support. Use monkeypatched histogram function to assert zero histogram calls for invalid RGB shape/dtype, and test both positive and negative axis indexing, 3-channel uint8/16 success, noncontiguous view, and mixed records. Benchmark rejected large FOV latency and peak memory but no fabricated speedup claims. Numba is not required.