Per-genome HTML/PDF report, written by detect by default

[hyphaltip]

Summary

• New matpredict report genome --run DIR [--pdf FILE]: one self-contained HTML report per detect run (no network or external files), printable to PDF.
  • Result first, in plain words: mating type and confidence; "Two idiomorphs found: needs review"; "No MAT locus called" with a next step; or "Not searched".
  • One card per locus: an inline-SVG gene-order figure (role, model status, strand, exons, alternate model, called-locus bracket, contig end), the idiomorph evidence and the gene table.
  • Then: what was searched, withheld candidate loci, provenance and a glossary.
• detect writes report.html by default. Opt out with --no-html or MATPREDICT_HTML=0; --pdf adds report.pdf. A report or PDF error is logged and never fails the run.
• Batch scripts set MATPREDICT_HTML=0: run_clade_panel.slurm, run_polish_ab.slurm, zygo_regression.py and run_holdout_benchmark.py. They use the variable rather than the flag because they may run an older frozen worktree, which would reject --no-html but ignores an unknown variable.
• detection_report.yaml gains a run provenance block. It records sample, organism, taxid, phylum, MATPredict version, database content SHA-256, taxonomy source, genome SHA-256/contigs/length/N50, parameters and timing. The block is additive, and older reports still render. detect gains --sample and --organism.
• WeasyPrint 69 joins the pixi environment and environment.yml. That is 25 conda packages and a 10.3 MB download, against about 300 MB for headless Chromium. No existing locked version changed. The Docker CI smoke test renders a PDF with no network. Headless Chrome is the fallback where WeasyPrint is missing.
• Design spec: docs/superpowers/specs/2026-10-07-genome-report-and-reads-intake-design.md. Part 1 (the report) is implemented here; Part 2 (reads intake for the on-demand service) is design only.

Review

Three rounds of an independent web- and data-design review (a briefed model agent, not a person) on rendered HTML, screenshots at 1200 px and 390 px in light and dark, and PDFs from both engines.

• Round 1: 6 P0 findings.
• Round 2: 1 P0, which the rewrite introduced: a hidden idiomorph margin that contradicted the table below it.
• Round 3: "high quality on screen in every case". The last print fixes were checked on the renders.

Details and the remaining limits: analysis/2026-10-07_report-design-review.md.

Test plan

• 36 new tests (tests/report/, tests/detect/test_provenance.py):
  • every fixture renders and shows every gene;
  • escaping of hostile sample and contig names (including the @page CSS string);
  • unique element ids and SVG colours that don't depend on page CSS;
  • margin wording, sample-name fallback, contig-end side and lane packing;
  • the default, flag and environment-variable switch;
  • a report error not failing detect, and a missing PDF engine keeping the HTML.
• Full suite on macOS: 1,097 passed. The 13 failures are identical to main on that machine (missing Linux binaries and the viz extras).
• PDFs rendered with WeasyPrint 70 and Chrome locally; 12 repeated Chrome renders, 0 failures.
• Docker CI build and the new offline PDF smoke test (no local Docker daemon, so this is the first check of WeasyPrint 69 in the image).
• Render a sample of real campaign reports on the HPCC (Basidiomycota, Mucoromycota, no-call); the fixtures other than Fola 50a are synthetic.

🤖 Generated with Claude Code