[Sad Catapella N3][Scientific provenance] Distinct percentile settings share the same serialized method identifier #35

Baseline and classification

Normalization Batch 3, pinned f784c95e3049483f87466150ffc5764375fa631d. Confirmed source-level settings identity collision, independently verified with isolated Python float formatting (not the package pytest suite). This is not the large-value arithmetic issue #32.

Evidence

• NormalizationSettings.method_name lines 139–145 constructs the only written percentile-method identifier with f"percentile {low:g}-{high:g}". Python's :g defaults to six significant digits.
• Parameters.from_columns lines 260–266 checks row/settings compatibility using only this method name, even though to_columns stores no exact percentile values or settings hash.
• Settings serialization does preserve the original percentile numeric values, so they may be materially different despite the identical method label.
• Tests check human-readable ordinary values, not collisions; fitting tests check a different method name, not two distinct settings whose names collide.

Deterministic counterexample

NormalizationSettings(percentile=(1.000001, 99.8)).method_name and NormalizationSettings(percentile=(1.000002, 99.8)).method_name both yield 'percentile 1-99.8'. A parameter row written with one is accepted by Parameters.from_columns(row, other_settings), despite distinct percentile thresholds. This is a proven identity collision; whether an actual dataset changes its fitted quantile values depends on distribution and pixel count.

Impact

Silently accepting stale/incompatible fitted normalization parameters, and ambiguous method/provenance reporting across runs. Especially problematic when preserving exact experimental parameter choices is a contract requirement.

Fix and acceptance

Treat the label as display text, not the full scientific settings identity. Store/validate exact method parameters or a canonical settings signature/version alongside parameters; maintain existing column layout only with explicit compatibility semantics. If changing the method string, version and migrate that contract and preserve old-row rejection behavior. Test high precision percentile pairs with identical short names, ordinary pairs, rgb, changed consistency, matching/contradictory settings, and round-trip exactly. No Numba role.