JPEG lab, slice 3: Quantize, Zigzag, Run-length, Huffman and Decode

[benjamin-small]

Stages 5–9 of the JPEG lab: Quantize, Zigzag, Run-length, Huffman and Decode. Plan: docs/superpowers/plans/2026-10-16-jpeg-lab-3-pipeline.md. The spec carries the Plan 3 amendment.

What's in it

• Quantize:
  • An editable Q table, applied through the rule config (tables). A "Custom Q tables · Standard" chip shows on every stage while custom tables are in use.
  • Coefficients before and after rounding, and the block the decoder rebuilds.
  • The R–D curve: q = 1…100 through the kit, within an ~8 ms frame budget, cached per picture and settings.
• Zigzag:
  • The viz animates the scan path and reports its progress through viz_readout().
  • The sequence strip marks the zero tail.
• Run-length: token chips that show each Huffman code and its amplitude bits. Together they spell out the block's bits exactly.
• Huffman:
  • Four tables from rule_inspect("huffman"), each with a histogram, a code tree and its cost against the entropy bound. The bound now compares code bits only; amplitude bits are shown separately.
  • The block's bits, coloured by part.
  • The file's segments, from rule_inspect("segments").
  • The size with the other Huffman mode.
• Decode:
  • A wipe between the decoded image and the original, and a brightened error map.
  • Artifact hints computed in Rust: blocking, ringing (none on pictures without hard edges) and colour bleed.
  • Re-save ×20 generation loss, disabled while custom tables are in use.

Tests

• cargo test --all-features: 85 + 311
• wasm-pack in headless Chrome: 47
• vitest: 526 in 37 files
• svelte-check: 0 errors
• fmt and clippy: clean (only the existing sierpinski warnings)
• A headless Playwright pass against the real engine on every new stage

Review

• Per-task reviews: two tasks needed one fix round each.
• Opus final review: "with fixes". One fix wave closed all 12 items, and the re-review found no new issues.
• Deferred minors are parked for Plan 4: mobile polish, the tab pattern, small test gaps.