parallel-fixture-audit

name	parallel-fixture-audit
description	Audit and fix gowkhtmltopdf implemented-CSS fixtures 60,61,62 (356 rules) by rendering all PDF pages to screenshots, scanning What it should do vs Effect (live) with 4-5 parallel sub-agents, collating defects, then fixing all broken items via parallel sub-agents. Use when user says screenshots of fixtures 60 61 62, 300+ rules, defect live, use 4-5 sub-agents, or runs /parallel-fixture-audit.

Parallel fixture audit — find and fix

Run the 60/61/62 implemented-CSS audit as a parallel find-then-fix loop. Do not run sub-agents sequentially.

Inputs

• Fixtures: testdata/golden/fixture-60-implemented-props-a.html (118), fixture-61-implemented-props-b.html (118), fixture-62-implemented-props-c.html (120) = 356 rules, catalog plans/0.2.6/catalog/implemented-fixture-split.json.
• PDFs: output/fixture-60-implemented-props-a.pdf, output/fixture-61-implemented-props-b.pdf, output/fixture-62-implemented-props-c.pdf (regenerate with --allow-local-files --font-path testdata/fonts/implemented-audit if stale).
• Screenshots: render every PDF page to PNG before scanning.

Phase 1 — Screenshots (all pages)

Render fresh screenshots, not cached:

python3 << 'PY'
import fitz
from pathlib import Path
srcs=[Path("output/fixture-60-implemented-props-a.pdf"),Path("output/fixture-61-implemented-props-b.pdf"),Path("output/fixture-62-implemented-props-c.pdf")]
out=Path("/tmp/fixtures_screenshots"); out.mkdir(exist_ok=True, parents=True)
dpi=150; m=fitz.Matrix(dpi/72, dpi/72)
for s in srcs:
    d=fitz.open(s)
    for i in range(d.page_count):
        pix=d[i].get_pixmap(matrix=m, colorspace=fitz.csRGB, alpha=False)
        pix.save((out/f"{s.stem}-p{i+1:02d}.png").as_posix())
PY

Expect 9 + 10 + 9 = 28 PNGs 1241x1754 at 150 DPI under /tmp/fixtures_screenshots/. Verify counts before scanning.

Phase 2 — Parallel find (4-5 agents, parallel not sequence)

Spawn 4-5 Task sub-agents in one turn (one Task call per message in same assistant turn). Each agent is read-only and owns a disjoint slice — never two agents on same package/file.

Suggested slices (cover all 356):

Agent	Slice	File ownership (read-only)
1	Flex / Align / Gap / Order / Place	flex layout, gap, place logic
2	Background / Border / Radius / Outline / Box-Shadow	paint, border_image, box_shadow, outline
3	Typography / Text / Font / Color	inline_paint, style_text_props
4	Layout / Sizing / Position / Overflow / Transform	layout, transform, overflow_clip
5	Grid / Multicol / Table / SVG / Misc	grid, multicol, layout_tables, pseudo_content

Each agent must:

1. Read its slice from the catalog and fixture HTML.
2. For each property trace: parse (style_cascade.go:normalizeVendorPrefix, style_properties.go:apply*) → ResolvedStyle (style.go) → layout/paint consumer (flex.go, layout_chrome.go, background_image.go, inline_paint.go, etc.) with file:line citations.
3. Compare Expected (What it should do column) vs Actual (Effect (live) cell and code reasoning) and mark pass / partial (fixture authorship) / broken (engine).
4. Return markdown table: Property | Fixture:Row | Expected | Actual | file:line | Severity.

Do not launch agents sequentially — emit all Task calls as parallel messages in one turn. Verify sub-agents produced turns before re-spawning.

Phase 3 — Collate

Merge the 5 tables. Prioritize broken (engine ignores value, paints wrong) over partial (fixture puts item prop on container, duplicate gap:20px+gap:4px last-wins, equal heights indistinguishable). Keep file:line citations. Report counts: total, pass, partial, broken.

Phase 4 — Parallel fix (4-5 agents, parallel not sequence)

Delegate fixes to 4-5 parallel Task fix agents, each with exclusive file ownership (no two agents edit same file). Do not run git commands unless user explicitly asks.

Example ownership:

• Agent 1: style_values.go, style_properties.go, style.go, style_cascade.go (place-content swap, max-height:%, grid line -1)
• Agent 2: inline_paint.go, style_text_props.go, inline_collect.go (wavy/dashed, emphasis, vertical-align, tab-size)
• Agent 3: layout.go, transform.go, filter.go, layout_chrome.go (translate:%, box-sizing min/max)
• Agent 4: overflow_clip.go, background_image.go, border_image.go, outline.go, style_advanced_props.go (overflow-clip-margin 1-4 values, bg clip, border-image width)
• Agent 5: grid.go, multicol.go, layout_tables.go, pseudo_content.go, style_paint_props.go (grid auto, caption-side logical, content:url, fill:none)

Each fix agent: read → minimal edit → go build ./internal/layout check → summary file:line. Keep file size < 2000 lines; extract rather than grow paint_flow.go.

Phase 5 — Verify and regenerate

1. go build ./internal/layout ./internal/convert
2. go test ./internal/layout -count=1 -short (note TestFixture60BackgroundImagesStayWithTheirRows is pre-existing spill, also fails on clean HEAD)
3. go test ./internal/convert -run TestGoldenCorpusAllFixtures -count=1 (update fixturePageBounds in internal/convert/golden_test.go if page count changes, e.g. 61 maxPages 9->11)
4. Regenerate output/fixture-60/61/62.pdf with --allow-local-files --font-path testdata/fonts/implemented-audit and re-render screenshots for next audit.

Phase 6 — Commit (only if user asks)

If user says commit/push and only Go (or Go + 60/61/62 HTML/PDF): stage exactly internal/layout/*.go, internal/convert/golden_test.go, and testdata/golden/fixture-60/61/62.html + output/fixture-60/61/62.pdf — not all output/*.pdf. No em dashes in messages.

Output shape

Report:

• screenshot counts and paths
• 5-agent defect tables with file:line
• collated priority fix list
• after-fix verification (build/test/golden) and regenerated PDF page counts

Do not write TODO.md or todos.json; use response todos.