fix: correct VP8 IDCT pass order in WebP encoder (chroma cast)

[ethanrous]

Summary

agno's WebP output had a chroma color cast (visible as a green/magenta tint) when decoded by libwebp — i.e. in browsers and the Weblens consumer. The root cause is in the VP8 inverse 4×4 DCT: it ran the row pass before the column pass, while the reference (vp8_short_idct4x4llm_c / libwebp TransformOne) runs the column pass first, then the row pass (with the +4 rounding bias).

The >>16 truncations in the trig terms make the two passes non-associative, so transposing them yields off-by-one reconstruction differences from any conformant decoder.

Why it was hidden

agno's encoder reconstruction loop and agno's own decoder shared the identical wrong order, so encode→decode round-trips through agno looked near-lossless. The defect only appeared when a conformant decoder (libwebp) decoded agno's output.

Why it presented as an accumulating chroma cast

• Chroma DC flows through this IDCT; luma DC goes through the order-independent Walsh–Hadamard transform (Y2). So luma was barely affected and the error was chroma-dominant.
• The ±1 chroma errors are not zero-mean on structured content, and they accumulate through intra-prediction feedback — growing from ~+2 U/V on a 1024px crop to ~+6 across a full frame. This matched all the observed clues (worst on low-chroma/hazy shots, error scaling with image size, crops decoding cleaner than the full image).

Fix

• Swap both idct4x4 implementations (codec/webp/transform.rs for the encoder's reconstruction, codec/webp/decode.rs for the decoder) to column-first / row-second, matching the reference.
• The GPU encoder path (webp_gpu.rs) only does RGB→YUV on the GPU and reuses the shared CPU encoder, so it is fixed automatically (no separate IDCT).
• Added a regression test (idct4x4_uses_vp8_reference_pass_order) pinning the pass order against the reference output for a block where the two orders differ. A round-trip test cannot catch this, because the encoder and decoder shared the deviation.

Test plan

• cargo test -p agno --no-default-features --features jpeg,png,webp --lib — 225/225 pass (incl. the new IDCT test)
• Decoded agno's WebP with libwebp (PIL/dwebp): chroma DC bias dV +6 → +0.09 on a full frame, +2 → ~0 on a crop; before/after on clean content shows the cast removed
• Verified the new test fails on the old (row-first) order and passes on the fix (the two orders differ on the chosen block)

Note: independent of the Sony ARW2 decode fix in #21.

[Copilot]

Pull request overview

Fixes a WebP chroma color cast by correcting the VP8 inverse 4×4 DCT pass order to match the reference implementation (column pass first, row pass second with +4 bias), ensuring agno’s encoder reconstruction matches conformant decoders (e.g., libwebp).

Changes:

• Swap IDCT pass order in the encoder-side transform implementation to column-first / row-second.
• Swap IDCT pass order in the decoder implementation to the same reference order.
• Add a targeted regression test that fails for the old (row-first) order and pins outputs to the reference order for a known-differing block.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
agno/src/codec/webp/transform.rs	Fixes encoder-side IDCT pass order and adds a regression test to pin reference behavior.
agno/src/codec/webp/decode.rs	Fixes decoder-side IDCT pass order to match the VP8 reference implementation.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.