gamut-avif: YCbCr matrix + 4:2:0/4:2:2 chroma subsampling on the lossy path

[justin13888]

• The lossy path codes GBR through the identity matrix at 4:4:4 — there is no colour decorrelation and no chroma subsampling at all.
• Verified at c2df11b:
  • crates/gamut-avif/src/encoder.rs — EncodeImage<Rgb8> calls Planar8::from_rgb8_identity_view(image) for both modes;
  • crates/gamut-av1/src/headers.rs:37-48 — Av1StillConfig pins chroma_subsampling_x/y = 0 and matrix_coefficients = 0 (Identity).
• This is the single largest RD gap in the crate, and it applies before any coding-tool comparison: libavif / SVT-AV1 / rav1e all encode 4:2:0 YCbCr by default.
• Deliver: BT.601/709 (and BT.2020) matrix coefficients plus 4:2:0/4:2:2 plane generation, selectable on AvifConfig, with matching av1C and colr signalling. Identity 4:4:4 stays the default for AvifMode::Lossless.
• Scope note: this sits between Optimize encoder half of gamut-avif for space-efficiency #171 (encoder heuristics) and gamut-avif: encoder support for high-bit-depth and non-Rgb8 inputs #251 (input/bit-depth/metadata surface, which explicitly defers lossy quality work to Optimize encoder half of gamut-avif for space-efficiency #171). Neither currently owns the output colour pipeline, hence a separate issue.
• Related: Implement space-efficient AVIF encoder #170, Optimize encoder half of gamut-avif for space-efficiency #171, gamut-avif: encoder support for high-bit-depth and non-Rgb8 inputs #251.

[justin13888]

Splitting this issue's two halves, because they are different orders of work.

Landed in #355 — the colour matrix, at 4:4:4:

• gamut_color::YcbcrMatrix — the H.273 §8.3 non-constant-luminance transform (BT.601 / BT.709 / BT.2020-NCL, full and studio range), derived from KR/KB into 16-bit fixed point; Planar8::from_rgb8_matrix.
• gamut_av1::Av1Colour + encode_still_intra_with, and the AV1 §5.5.2 color_config() non-shortcut branch (the explicit color_range bit).
• AvifConfig::{matrix, range} via AvifEncoder::{with_matrix, with_color_range}, mirrored into av1C and colr. AvifMode::Lossy now defaults to BT.709; Lossless stays identity.
• Backend seam: colour on Av1EncodeRequest, validated against the returned stream's parsed color_config().
• Decode parity: the RGBA surface reads BT.709 and BT.2020-NCL.

Measured 31–38% smaller at equal quality on correlated photographic content (19% on the small golden fixture); byte-exact against libaom and dav1d at both ranges.

Still open here — 4:2:0/4:2:2 chroma subsampling. This is gamut-avif's M2 milestone, not a bolt-on, because gamut-av1 is 4:4:4 throughout:

• Planar8 carries one width/height, and FrameEncoder one coded_w/coded_h for all three planes (tile.rs:439-465), read through a single-stride sample() (tile.rs:640).
• Chroma contexts (set_ctx, txb_skip_ctx, dc_sign_ctx) are bound by the luma mi_cols/mi_rows; the residual loop steps chroma over the luma block extent (tile.rs:1461-1508).
• CfL reads luma unsubsampled (tile.rs:2397); §7.11.5 needs the 2×2 (420) / 2×1 (422) average.
• Deblock derives the chroma transform size from the luma MI grid and CDEF filters chroma on the luma 8×8 grid with an identity Cdef_Uv_Dir (filter.rs:213,543).
• New spec logic with no home yet: HasChroma / get_plane_residual_size (§5.11.5/§5.11.38) — with 4:2:0 a 4×4 luma block does not code its own chroma, which interacts with uv_mode, cfl_alpha signalling and the coefficient contexts. The encoder does emit sub-8×8 blocks today.
• Signalling: seq_profile 0 (4:2:0) / 2 (4:2:2), chroma_sample_position, and compatible_brands — MA1B for 4:2:0, while 4:2:2 (profile 2) matches neither AVIF profile brand.
• Oracles: tooling/libavif-oracle's decode_avif hard-rejects non-4:4:4 and would need relaxing; the ten pinned FNV-1a bitstream checksums and both AVIF goldens sit downstream.

Suggested sequencing once picked up: per-plane geometry refactor first (mechanical, no behaviour change), then 4:2:0 and 4:2:2 together (they share HasChroma), then the filters and CfL, then brand/profile selection.