Hill-climb layout quality to beat mermaid

Incremental layout algorithm improvements where each fix is validated against baseline + other fixtures (HYBRID gate)

• · kymo flowchart corpus (target fixture)
• · baseline layout quality scores (kComposite, mComposite)
• · hypothesis: one kymo-layout code change per round

• · git diff per accepted fix
• · layout quality scores after each round
• · final leaderboard (kymo vs mermaid.js)
• · unified diffs ready for PR

• · git worktree (isolation per hypothesis)
• · wasm-pack (rebuild kymo-mermaid wasm)
• · benches/layout-algorithms (scoring harness)
• · pytest, cargo test (HYBRID gate verification)
• · Chrome browser

• · Google Chrome at default path (or $CHROME override)
• · wasm32 target installed; wasm-pack available
• · benches/layout-algorithms/ npm ci completed
• · kymo-mermaid wasm built and cached
• · baseline layout-quality scores computed

• · quality metric regression on other fixtures
• · pixel delta vs mermaid.js worsens (visual fidelity loss)
• · golden layout tests fail (pytest)
• · cargo test failures in kymo-layout
• · wasm rebuild fails or mismatches expectations

• · HYBRID gate: must rise on target AND pass 4 checks (pixel, golden tests, cargo test, no regression on other fixtures)
• · Worktree isolation: never touches main branch
• · Baseline/leaderboard published before round 1
• · EPS slack (0.5 composite units, 0.3 pixel ratio) gates acceptable drift
• · Max 2 consecutive dry rounds before halt
• · Proposed diffs for human review (never auto-merges)