[jcoyne]

Fixes #321.

Problem

A three-component JP2 tagged with an embedded IEC 61966-2-1 sRGB ICC profile failed decode_region_into and decode_region_scaled_pow2_into to Rgb8/Rgba8:

unsupported: backend color space cannot be mapped to requested 8-bit pixel format

The only working route was decode_srgb8, which decodes the whole image. That isn't practical for large IIIF sources.

Change

The 8-bit output paths now treat a three-channel ICC profile as plain RGB for Rgb8/Rgba8 output when the profile is equivalent to sRGB. Samples are written without a colour transform. A profile qualifies when:

• it is a restricted matrix/TRC RGB profile, the same kind decode_srgb8 accepts;
• its linear RGB-to-sRGB matrix (moxcms transform_matrix) is within 2.5e-4 (off-diagonal row sums) and 1e-3 (diagonal) of identity. Near zero, one 8-bit sRGB code value is about 3.0e-4 of linear light, so the off-diagonal limit keeps crosstalk under one code value;
• its 8-bit transform to sRGB moves no red, green, blue or neutral ramp sample by more than one code value.

The ramps alone are not enough. A negative off-diagonal term clips to zero on a single-channel ramp and cancels on neutrals, but it still shifts mixed colours: a test profile with matrix [[1.01, -0.01, 0], [0, 1, 0], [0, 0, 1]] moved (30, 255, 0) to a red of 10. That is why the matrix is checked directly.

This applies to decode_region_into, decode_region_scaled_pow2_into, the prepared region path, the whole-image decode_into fallback, and the streamed stripes of decode_rows_u8_bounded. The colour space is resolved once when each row sink is built, so the profile isn't re-checked per band. Other ICC profiles and 16-bit output stay unsupported, with the same error.

Tests

New file crates/j2k/tests/srgb_icc_region.rs:

• With an sRGB ICC profile, output is byte-identical to the same image tagged with the enumerated sRGB colour space. Checked for region Rgb8/Rgba8, scaled regions at 1 and 2 levels, and whole-image decode_into, for both moxcms's parametric sRGB profile and a 1024-entry sampled-curve sRGB profile like the common HP one. Whole-image output is also within one code value of decode_srgb8.
• Display P3, Adobe RGB and ProPhoto profiles are still rejected.
• decode_rows_u8_bounded at 1, 5 and 16 rows per stripe matches the enumerated-sRGB image. This test fails without the stripe-sink change.
• The crosstalk profile above is rejected. This test fails without the matrix check.

The three sRGB tests fail without the source change. cargo test -p j2k passes. A manual check against the macOS system profiles accepted sRGB Profile.icc and rejected Display P3, ITU-709, Generic RGB and AdobeRGB1998.

Notes

• The profile is parsed and checked on every 8-bit decode call of an ICC-tagged image; the result isn't cached on the decoder.
• The matrix and ramp limits are each bounded separately. A profile right at both limits could in theory be off by a little more than one code value; real sRGB profiles are far inside them.

[jcwal1516]

lgtm, thanks for this!