salmon 1.12.0 — final C++ release

salmon 1.12.0 — release notes

This is the final release of the C++ implementation of salmon. Future
development continues in salmon 2.0, a from-scratch Rust rewrite that is
faster, easier to build and install, and that fixes the issues below natively.
The 1.12.0 line will remain available for reproducibility and emergency fixes
(see "Continuity" below), but new features and performance work land in 2.0.

1.12.0 is a correctness-focused release. Every fix here was discovered while
building and cross-validating the Rust rewrite against C++ salmon, and then
backported to C++ so existing 1.11.x users benefit immediately. Indices built
with salmon 1.11.x are compatible — no reindexing is required for 1.12.0.

---

Correctness fixes

1. Selective alignment: k-mer orientation in the SSHash streaming lookup

The SSHash streaming k-mer lookup reused the canonical-relative orientation
formula for a query-relative result, flipping hitFW for non-canonical query
k-mers. Reads whose seed support reduced to a single k-mer (e.g. short reads with
one sequencing error) could be mis-placed on the wrong strand and discarded.

• Affected versions: SSHash-based releases only (salmon ≥ 1.11.0).
• pufferfish pinned to the streaming getRefPos orientation fix (5dce7f4).

2. --seqBias: observed sequence-bias model was never trained

Fix: numBiasSamples was zeroed in the pipeline context before option parsing applied
its default, so the bias-sampling gate never fired. The observed ("foreground")
model stayed uniform and the correction degenerated to 1 / P_expected(context),
inflating effective lengths (often beyond transcript length). Fixed for both
the reads path and the alignment-based (-a) path.

• Affected versions: salmon ≥ 1.11.0 only.
• Impact: yeast ERR458493 — observed-bias samples collected 0 → ~430k;
  transcripts with effLen > length 4,015/6,571 → 117/6,571; bias-corrected
  effective lengths now match an independent re-implementation at Pearson
  0.91 → 0.99943. --numBiasSamples now also takes effect (it previously did
  not). Clean non-bias runs are unaffected.

3. Alignment mode: pair mates by mate fields, not stream adjacency

Improvement: This was not a correctness bug, but was a laxity in assuming the precise structure of the input.
BAMQueue assumed paired-end records are adjacent in the BAM stream. Aligners
that emit all R1 records then all R2 records within a name-collated group
(e.g. bowtie2 -k/-a) violated this, causing salmon to pair read1-with-read1
(same-orientation) and silently drop fragments. Mates are now paired via
reciprocal RNEXT/PNEXT (disambiguated by the HI tag) using a small
read-ahead interleave buffer.

4. Positional bias: additive-smoothed factor + 3′ coordinate fix

Improvement: Corrected the positional-bias factor to use additive smoothing and fixed the 3′
coordinate computation used when applying the positional model.

5. --writeMappings: always flush SAM records to disk

Fix: The SAM/BAM sink could retain buffered records at shutdown; output is now
force-flushed so all mapping records are written.

6. BAMQueue: do not pre-allocate a 2M-object pool at startup

Improvement: Removed an unconditional ~2M-object pool allocation at process start, reducing
startup memory and latency for alignment-mode runs.

---

Continuity

• salmon 1.12.0 is the last C++ release. Development moves to salmon 2.0
  (Rust): a single portable binary, cargo/conda installation, the same
  quant.sf output (drop-in for tximport), compatible inferential-replicate
  output (bootstrap/Gibbs), and a new alignment-free --sketch mode.
• The C++ source is preserved on the cpp branch; a salmon-cpp conda package
  may be provided for users who need the C++ line for reproducibility or
  emergency fixes.
• salmon 2.0 indices must be rebuilt (new index format); 1.12.0 indices are
  unchanged from 1.11.x.

Acknowledgements

These fixes came out of the salmon Rust rewrite effort and the cross-validation
between the two implementations.

---

Full Changelog: v1.11.4...v1.12.0