[joepetjr]

Summary

• LUT-1. load_cube() and Cube now raise CubeError on NaN, infinity, or a value that overflows float32, in the domain, the 3D table or the 1D curves. A directly constructed Cube with an inverted domain is rejected too. Before, these values reached the final cast as undefined pixels.
• LUT-2. apply_cube() accepts only uint8 and uint16 pixels. Any other dtype raises CubeError. Before, every dtype that was not uint16 was divided by 255, so a float image in [0, 1] was silently crushed.
• README and CHANGELOG updated.

Plan units: LUT-1, LUT-2 (portfolio remediation 2026-09-29).

Behavior change to review

A caller that passed float or signed arrays to apply_cube() used to get a wrong image. It now gets an error that names the dtype and says to convert. No normalized-float API is added.

Finite table values outside [0, 1] are still accepted, as the format allows.

Test plan

• LUT-1: nan, inf, -inf, 1e999 and 1e39 in table, domain and 1D curve; direct construction. Seventeen cases were red before the change (DID NOT RAISE CubeError).
• LUT-2: nine unsupported dtypes plus a float [0, 1] image. Ten were red before the change.
• Valid 1D and 3D fixtures apply with warnings treated as errors and keep their known output.
• Full suite locally: 71 passed. ruff check . clean.

Acceptance

• A NaN cube fails before application: test_non_finite_table_entry_raises_before_apply, test_non_finite_domain_raises, test_non_finite_1d_entry_raises.
• Valid fixtures keep known output with no cast warnings: test_valid_fixtures_apply_without_runtime_warnings.
• Float and signed arrays cannot silently use 255 scaling: test_unsupported_pixel_dtype_is_rejected_not_scaled.
• Supported dtypes keep precision: test_uint16_precision_is_kept_beyond_8_bits, test_uint8_known_output_is_unchanged.