outFilterMismatchNoverLmax is applied to the read length, not the mapped length

ReadAlign_mappedFilter.cpp applies outFilterMismatchNoverLmax to the mapped length of the alignment, while outFilterMismatchNoverReadLmax is the one applied to the read length.

In src/align/read_align.rs both are currently divided by read_seq.len():

let read_length = read_seq.len() as f64;
...
let mismatch_rate = t.n_mismatch as f64 / read_length;
if mismatch_rate > params.out_filter_mismatch_nover_lmax { ... }

Consequence: for a soft-clipped alignment the denominator is too large, so the filter is more permissive than STAR's. A 100 bp read mapping 50 bp with 12 mismatches gives 0.12 here (kept at the 0.3 default) against STAR's 0.24 (also kept), but at 20 mismatches over 50 mapped bases it is 0.20 here and 0.40 in STAR, which STAR rejects and this does not.

The fix is to divide by the summed mapped length of the alignment's exons. It is deliberately not bundled into the PR that adds outFilterMismatchNoverReadLmax, because changing the denominator moves the yeast benchmark and wants its own before/after measurement.

Until then the two parameters share a denominator, so outFilterMismatchNoverLmax and outFilterMismatchNoverReadLmax behave identically.

Start in src/align/read_align.rs, where both mismatch rates currently use read_seq.len(), and inspect how the alignment’s exons expose their mapped lengths. Update the mapped-length calculation for outFilterMismatchNoverLmax, then run the relevant tests and compare the yeast benchmark before and after to confirm the denominator change.