ci: opt-in large-dataset differential against STAR

[BenjaminDEMAILLE]

Closes #244.

Why

The suites in tests/ use a synthetic micro-genome and finish in seconds. That is why the two bugs reported from nf-core/rnaseq data (#31, #48) were invisible to them: both are about aggregate behaviour over tens of thousands of reads, not about one alignment.

What runs

test/nfcore_diff.py fetches the nf-core/rnaseq test dataset from public URLs (50 000 paired reads, S. cerevisiae chrI plus the GFP transgene — the same fixture the reporter used), builds both indexes, runs STAR and rustar-aligner, and compares:

• uniquely mapped, multi-mapped, unmapped: too short, unmapped: other, each as a fraction of input reads;
• the deepest multimapper (max NH), as a ratio.

Thresholds are the measured gaps plus headroom, so today's state passes and a regression trips them. --report-only prints the comparison and exits 0, which is what to use while a difference is being investigated.

Checked against a known regression

Run on main right now, the script fails on exactly the two buckets #247 fixes, and reports #31's depth gap:

metric                                         STAR     rustar      delta
Uniquely mapped reads number                  41691      41665    0.052%
Number of reads mapped to multiple loci         934        961    0.054%
Number of reads unmapped: too short            3766       7374    7.216%
Number of reads unmapped: other                3609          0    7.218%
deepest multimapper (NH)                         14         20      1.43x

FAILED:
  Number of reads unmapped: too short: 7.216% > 1.000%
  Number of reads unmapped: other: 7.218% > 1.000%

With #247 merged those two lines read 3778 and 3596, inside the thresholds.

How it is triggered

Opt-in, never on every push: manual dispatch, or the large-tests label on a pull request. The dispatch form takes a runner label, so it can be aimed at a scverse AWS runner without editing the workflow — that is the part of the issue I cannot do from here, since it needs the runner group to exist.

STAR comes from Ubuntu's rna-star package; building it from source would triple the job runtime without changing what is measured.

One note carried into the script: --readFilesCommand zcat yields zero reads on macOS for both aligners, silently, so the fixture is decompressed up front rather than piped.

🤖 Generated with Claude Code