Correctness and performance pass: mapping positions, EM convergence, Gibbs prior, SAM, GFF3, CLI

[BenjaminDEMAILLE]

Rebased onto current master. Changes that master already covers (decoy block located by name, deterministic eq-class order, replicate scaling, parallel digamma and balanced EM shards, allocation-free alignment cache, posterior reuse) were dropped. One commit per change.

Correctness

• Mapping positions: read and fragment positions now come from the projected read ends (like pufferfish's approxReadStartPos), not from the first/last seed. Before, a read whose first or last k-mer carried a mismatch got a shortened fragment length, which biased the FLD, library-format and dovetail detection, the SAM POS and the bias coordinates. This also fixes an off-by-one in the 5' position of reverse single-end reads and orphans.
• Recovered orphan pairs: ref_pos, the partner's position and the positional-bias ends now describe the fragment, as for a concordant pair. recover_mate clamps its window to the transcript (no negative window start).
• Paired seq-bias (ISR and similar): the 5'/3' contexts were attributed to the wrong end when mate 1 is the reverse mate.
• Refseq store: reference lengths are checked against the index (duplicate names no longer write a corrupt store).
• EM/VBEM: with no transcript above the convergence cutoff (e.g. no fragments), the loop ran to max_iter. It now counts that case as converged, as salmon does, and filters on alpha_out. Applied to the plain, SQUAREM and DAAREM loops.
• Gibbs prior under --perNucleotidePrior: uses vbPrior per nucleotide (matching the VBEM point estimate) instead of max(1, vbPrior), and only with VBEM.
• Ambiguity counts (ambig_info.tsv): accumulated in u64.
• SAM/BAM: CIGARs no longer contain zero-length 0M ops when a read ends exactly at the transcript start or starts exactly at its end.
• Bias convolutions: fld_low == 0 no longer indexes position -1; the FFT path uses the same fragment lengths as the scalar loop.
• -g with Ensembl GFF3: transcripts without gene_id are resolved to genes via Parent.
• CLI: --fldMean/--fldSD (finite, positive), --numGCBins/--conditionalGCBins (1..=101), --thinningFactor (>= 1) and the scoring flags are validated at parse time; salmon's negative --mp form is accepted.

Performance

• Gibbs: multinomial buffers are reused across rounds instead of allocated per class.
• FLD add_val: the shared sum/tot_mass atomics are updated once per fragment instead of once per kernel entry.

Validation

• Every commit builds (cargo check --workspace --all-targets).
• cargo test --workspace --release, cargo fmt --check and cargo clippy --workspace --all-targets -- -D warnings pass on the last commit.
• New regression tests: projected fragment length, empty EM convergence, parallel vs sequential EM, CIGAR overhangs, Ensembl GFF3.

🤖 Generated with Claude Code