feat: BED-interval rRNA quantification (split_bam) and featureCounts -M/-O

[BenjaminDEMAILLE]

Closes #111.

Addresses all three asks in the issue.

1. split_bam: BED-interval classification

New module src/rna/rseqc/split_bam.rs, the counting side of RSeQC's split_bam.py. Runs inside the existing single pass, disabled unless a BED file is supplied:

rustqc rna sample.bam --gtf genes.gtf --rrna-bed GRCh38_rRNA.bed

Output rseqc/split_bam/{sample}.split_bam.tsv:

category	count	percent_of_total	percent_of_usable
in	107	21.9262	22.1532
ex	376	77.0492	77.8468
junk	5	1.0246	NA
total	488	100.0000	NA

Classification follows split_bam.py: unmapped/QC-failed records are junk; otherwise the read's own start, and the mate's start when the mate is mapped, are tested against the intervals. Secondary and supplementary alignments are deliberately not filtered — that is exactly what lets this route see the rDNA multi-mappers that featureCounts' default single-hit rule discards.

Not implemented: the three output BAM files (.in.bam / .ex.bam / .junk.bam) that split_bam.py writes. Writing three extra BAM copies would defeat the single-pass design; the docs point at samtools view -L for that. Happy to add them behind a flag if you'd rather have full parity.

2. featureCounts -M / -O

-M, --count-multi-mapping and -O, --count-multi-overlapping (plus rna.featurecounts.count_multi_mapping / count_multi_overlapping), applied to the gene-level and biotype-level featureCounts classification. dupRadar's duplicate-rate matrix is untouched — it tracks multi-mappers separately by design.

On tests/data/test.bam, -M -O moves the 10 multi-mapping reads from Unassigned_MultiMapping into Assigned (473 → 483).

3. Docs note

A caution box on the featureCounts page spelling out why biotype %rRNA reads near zero on stock GRCh38/GENCODE (45S rDNA unannotated + multi-mapper drop), pointing to -M/-O and to split_bam. Plus a split_bam section on the RSeQC page and CLI/config reference entries.

Verification

• samtools view -c tests/data/test.bam chr1:1001-3000 = 107, matching the in count for a chr1 1000 3000 BED
• New unit tests: BED parsing (half-open coordinates, malformed lines, empty file), classification of a record mix including a secondary alignment and a mate-side hit, and -M / -O behaviour at gene and biotype level
• cargo test — 207 lib + 18 integration tests pass
• cargo fmt --check, cargo clippy -- -D warnings clean

🤖 Generated with Claude Code