[BenjaminDEMAILLE]

Part of #354 (finding 20).

Stacked on #346. perf/hotpaths is on a fork, so GitHub cannot use it as the base of a PR here; the base is main, and until #346 lands the diff shows its commits too. Only the last commit (bcddd31) is this PR: review with git diff aeba641 bcddd31. Retarget or rebase after #346 merges.

Problem

The tree carried code nothing runs. One block of it cost time: the stitcher scored every seed before extension for every window, then discarded the result with let _ = best_pre_score. The stdout writers were copies of the file writers that had drifted: SamStdoutWriter lacked the [SAM-MISMATCH] check.

Change

12 files, 265 insertions, 1056 deletions. Each item was grepped across src, tests, examples, benches and scripts, then confirmed by the compiler once removed.

Removed	How it was verified dead
Stitcher block ending in let _ = best_pre_score, and the pre_ext_score field it filled	Result discarded; no other reader of the field
Reverse-strand branches of the stitcher	stitch_seeds_core returns stitch_cluster with is_reverse: false, the only cluster stitch_recurse, stitch_align_in_place and finalize_transcript receive. is_reverse goes from extend_alignment(_cached), score_region, count_mismatches_in_region
Reverse-strand branches of the junction scan (not named in the finding, same cause)	find_best_junction_position_cached is called only from stitch.rs, always forward. is_reverse, n_genome, genome_offset go from find_best_junction_position(_cached) and ScanArgs; score_gap_with_strand merged into score_gap. examples/jrbench.rs passed these arguments and is updated
BufferedBamRecords, SamWriter::write_alignment and its test	No production caller. test_sam_output_includes_rg_header_and_tag used write_alignment; it now uses build_unmapped_record and write_batch, and still checks the RG header and tag
Seed::find_paired_seeds, Seed::get_genome_positions, combined_transcript_for_projection	Called only by their own tests, which go with them
parse_gtf, extract_junctions_from_exons	Wrappers used only by tests, which now call parse_gtf_configured and extract_junctions_configured
word_comp_length	No caller
Duplicate in src/index/mod.rs	The transcript tables and sjdb files were written by two copies; both paths now call write_transcript_tables and write_sjdb_files
SamStdoutWriter, BamStdoutWriter, SortedBamStdoutWriter	Replaced by SamWriter<W>, BamWriter<W> (default type parameter is the file writer, so existing call sites are unchanged) and SortedBamWriter::create_stdout

Every item was dead in production; none had to be kept. stitch_is_reverse and original_is_reverse, which carry the read's real strand, are untouched.

These were pub items of the library crate, so this is an API removal, marked in CHANGELOG.md.

One intended difference: --outStd SAM now has the file writer's [SAM-MISMATCH] check. No run below triggers that check, so the check itself on stdout is not exercised; the output of those runs is identical.

Conflict with #355

git merge-tree reports conflicts with fix/output-completeness (#355) in CHANGELOG.md, src/io/sam.rs and src/lib.rs. This branch does not add finish, because #355 has not landed and this PR changes no behaviour. Whichever lands second must keep the flush: #355's SamWriter::finish goes on the generic impl<W> SamWriter<W>, the AlignmentWriter impl on impl<W> AlignmentWriter for SamWriter<W>, and its SamStdoutWriter::finish is dropped along with that type. I did not carry out or test this resolution.

Measured

macOS, aarch64, 16 cores. Baseline: clean build of aeba641. Comparison: bcddd31.

Output. 84 scenarios, each at --runThreadN 1, 4 and 8, every binary under the same argv[0] and relative prefix so that command lines in headers match. md5 of every output file, raw BAM and SAM included, plus decoded BAM records and Log.final.out without its clock lines. Log.out, Log.progress.out and stderr are not in the md5s.

Set	Scenarios	Rows compared	Rows differing
Generated fixture (20 000 reads) and tiny fixtures	75	2367	4 (below)
Earlier fixtures, including rejected inputs	9	132	0

The 4 rows are Solo.out/{Gene,GeneFull}/CellReads.stats at 4 and 8 threads in one scenario. This is existing behaviour: two runs of the baseline differ from each other in the same file, with the same md5 once sorted (64136fbd), so only the row order changes between runs. At 1 thread the file is identical. Not touched here.

Mode	Scenarios cover
Single-end	SAM, BAM unsorted and sorted, --outStd in all three forms, Within, TranscriptomeSAM, GeneCounts, Fastx, chimeric junctions and WithinBAM, BySJout, fixed clip, clip to empty, adapter, CellRanger4, bedGraph, WASP, splicing tag, per-file RG, NoQS, None, 2-pass, read limit
Paired-end	The same where it applies, plus half-mapped pairs
Solo	SAM, BAM, Gene, GeneFull, SJ, Velocyto, EM, CB_samTagOut, clip, CR4, RG, None
Solo paired	Third barcode file and --soloBarcodeMate 1, half-mapped, CB_samTagOut, clip, RG, None

The fixture was checked to reach these paths on the baseline: SE 15 965 unique, 1614 multi, 1417 too short, 1003 other, of 20 000 reads; a second fixture gives 2151 (SE) and 2152 (PE) chimeric reads, with as many supplementary records under WithinBAM; relaxed --outFilter*OverLread gives 92 records of half-mapped pairs in PE and in solo paired.

Scenarios the program refuses: sp_tag_m (one invalid parameter combination of mine, all 3 thread counts) and 5 of the 9 earlier scenarios (input files the reader rejects; the 3 I compared give the same error text). Same exit code 1 on both binaries; they prove little beyond that.

Exit codes, decoded records and non-header bytes are also identical across 1, 4 and 8 threads within each binary, apart from the same CellReads.stats rows.

Speed. --runThreadN 8. The machine was not idle: unrelated bwa-mem2 and salmon jobs were running, so wall time is noisy (the same binary varies between 4 and 16 s), and I compare instructions retired and CPU time. From /usr/bin/time -l, binaries alternated.

Case	Input	Instructions (median), base	this PR	Change	CPU s (median of 5), base	this PR
SE, largest fixture	B200.fq, 801 000 reads, 176 MB	103.47 G	93.32 G	−9.8 %	10.94	10.29
SE	800 000 reads	376.35 G	344.33 G	−8.5 %	30.48	28.67
PE	800 000 pairs	1466.2 G	1400.2 G	−4.5 %	110.70	106.69
Solo	800 000 reads	382.25 G	350.08 G	−8.4 %	30.34	28.52
Solo paired	800 000 pairs	1472.7 G	1407.1 G	−4.5 %	111.88	106.33

Instructions: 3 runs for SE and solo, 1 run for PE and solo paired. The reads are synthetic and short; expect a different figure on real data.

Tests

cargo test --release passes (740 library tests and all integration tests, tests/determinism.rs and tests/phase9_threading.rs included). cargo clippy --all-targets has no warnings; cargo fmt applied. No test was added: this removes code.

Not checked

• I did not run test/parity_suite.sh, and did not compare against STAR here; the claim is "same output as aeba641".
• Real sequencing data, gzip input, Linux.
• docs-old still names pre_ext_score and get_genome_positions; it is historic and left alone.

Also dead but not named in the finding, left in place: AlignmentScorer::score_gap is called only by tests, and stitch_seeds / stitch_seeds_with_jdb have no production caller beyond a re-export and each other.

🤖 Generated with Claude Code