fix(stats): paired-end unmapped reads split between too short and other

[BenjaminDEMAILLE]

Closes #48, reproduced on the reporter's own fixture.

The rule

STAR decides the bucket in ReadAlign_mappedFilter.cpp:

if ( nW==0 ) {//no good windows
    statsRA.unmappedOther++;
    unmapType=0;
} else if ( (trBest->maxScore < P.outFilterScoreMin) || ... ) {//too short
    statsRA.unmappedShort++;
    unmapType=1;

So other means no window was found at all, and too short means a window existed and its best transcript failed the score or length thresholds.

The single-end path here already made that distinction. The paired-end path returned TooShort for every unmapped pair, which is why other was permanently zero.

Measured on the reported fixture

nf-core/rnaseq test data: 50 000 pairs, S. cerevisiae chrI + GFP transgene, --outFilterMultimapNmax 20, otherwise defaults.

	too short	other
before	7374	0
after	3778	3596
STAR 2.7.11b	3766	3609

The totals already agreed (7374 against STAR's 7375), which is what made this a labelling bug rather than a counting one. The remaining 12-13 reads at the boundary are a separate question: they are reads where the two implementations disagree about whether a window exists at all, not about how to label one.

Tests

tests/unmapped_buckets.rs plants all three populations: pairs that map, pairs that seed but cannot clear the score threshold (a window exists, so "too short"), and pairs with no genomic seed at all ("other"). It also asserts the buckets sum to the input count.

Checked against a control: with the fix reverted, the paired-end test fails and the single-end one still passes, which is exactly the shape the bug had.

cargo test, cargo clippy --all-targets -- -D warnings on a cold cache, and cargo fmt --check are green.

Note on the reproduction

--readFilesCommand zcat produces zero reads on macOS for both STAR and rustar-aligner, silently: the run succeeds and Log.final.out reports 0 input reads. I hit it before getting to the real comparison. Unrelated to this fix, but worth knowing when reproducing anything from a gzipped fixture on a Mac; #239 makes --readFilesCommand gunzip -c work by running the command through a shell, which sidesteps it.

🤖 Generated with Claude Code