v2.11.0

Added

• bamtocounts: --exclude-multimapping drops reads with an NH tag > 1,
  matching the default multi-mapping filter of featureCounts (no effect if the
  aligner does not write NH).

Fixed

• bamcountrefs: memory no longer grows with references × samples × metrics.
  Output tables are streamed row by row instead of being built in memory, and
  per-sample results are freed as soon as they are merged. On a metagenome
  co-assembly (3.59 M contigs × 21 BAMs, --all-metrics) the previous build
  was killed at over 60 GB; it now completes with a 30–35 GB peak, and on 4
  BAMs it went from 107 s / 17 GB to 64 s / 10 GB. Output is byte-identical.
• bamcountrefs: -W/--workers is now honoured. It was parsed but ignored,
  and one thread was started for every input BAM at once.
• bamcountrefs: the RPKM denominator is now the number of reads passing the
  filters (-F, -Q, -P). It was the BAM index "mapped" total, which
  includes supplementary, secondary, duplicate and QC-fail alignments, so RPKM
  was too low whenever filters removed reads (~1.3% on a minimap2 co-assembly
  BAM). Values now match coverm.

Changed

• bamcountrefs: RPKM values change for any BAM where the filters remove
  reads (see above). TPM and all other metrics are unchanged.
• bamcountrefs: with many BAMs and a low -W, fewer files are processed in
  parallel than before, since the old build ignored -W. Raise -W to trade
  memory for parallelism.