[Sad Catapella N6][Contract consistency] RGB Parameters.select permits invalid channel subsets that cannot round-trip through from_columns #50

Baseline and classification

Sad Catapella normalization Batch 6 at f784c95e3049483f87466150ffc5764375fa631d. Confirmed public-API inconsistency by source inspection; no repository test executed. This is distinct from #38 (fractional channel coercion) and #23 (allowed RGB dtypes).

Evidence

• Parameters.select() lines 132–159 accepts a subset of any stored channel indices, without distinguishing RGB from fluorescence. A fitted three-channel rgb percentile 1-99.8 parameter set can become a one-channel RGB parameter object with channels=(2,).
• Parameters.to_columns() lines 209–226 emits only the selected channel's offset and scale; Parameters.from_columns() lines 301–310 calls _check_rgb_columns() lines 528–548, which requires exactly C0, C1 and C2, so from_columns(selected.to_columns(), rgb_settings) fails.
• The documented generic select contract describes splitting channels. But CellPatcher's confirmed RGB workflow explicitly rejects RGB split_channels before patching, and tests this rule, so RGB subset support is not established as intended.
• tests/test_rgb.py lines 191–220 cover full RGB round trips and malformed subset columns, but never call select on RGB parameters.

Minimal deterministic repro

p = Parameters('rgb percentile 1-99.8','image','g',True,(1.,)*3,(4.,)*3); p.select([2]) currently returns an object that applies to one grayscale channel and serializes, but Parameters.from_columns(p.select([2]).to_columns(), NormalizationSettings(image_type='rgb')) raises an RGB column validation error.

Scientific impact / decision

RGB relies on one jointly fitted RGB normalization: leaving one RGB channel separated can be interpreted as a legitimate independent RGB patch, contrary to the consumer contract. Silent accept at one boundary and refusal at the next produces inconsistent run behavior.

Fix and acceptance

Prefer rejecting Parameters.select with a proper RGB-specific error when a strict subset of (0,1,2) is requested; allow selecting all channels unchanged. If the consortium explicitly wants post-fit RGB channel extraction, instead add a versioned, documented provenance distinction so serialization and consumers support it consistently—never silently coerce a subset back to a full RGB group. Add direct select, apply, to/from-columns, all-channel selection, fluorescence subset, CellPatcher cross-brick regression tests. Do not modify the shared-channel percentile algorithm. No Numba role.