[BenjaminDEMAILLE]

Part of scverse#354 (finding 3).

This PR is on the fork because perf/hotpaths exists only here, and a PR on scverse/rustar-aligner cannot use it as a base. It can move there once scverse#346's branch is merged or copied.

Stacked on scverse#346 (perf/hotpaths, tip aeba641). Only the one commit is this PR. It merges without a conflict with the PR for finding 2, so neither is stacked on the other. Pairs only reach zero differences with both, see the tables.

Problem

--alignEndsType EndToEnd left the inner ends of a pair's mates soft-clipped. With 4 mismatches at the 3' end of each mate:

	CIGARs	AS
STAR	100M / 100M	182
perf/hotpaths	96M4S / 4S96M	190
This branch	100M / 100M	182

The end-to-end decision was made in finalize_transcript, on one mate's slice, after the pair had been chosen and scored. Its mismatches never counted in the pair's score or limits, and a mate that could not reach its end could still be clipped into a pair. Extend5pOfRead1, Extend5pOfReads12 and Extend3pOfRead1 went through the same code.

What STAR does

Observed output of Homebrew STAR 2.7.11b; source not read.

• Both orientations of the pair, and mismatches at the outer ends instead, give 100M / 100M, AS:182, nM:8.
• Extend3pOfRead1 extends only read 1's 3' end: 100M and 4S96M, AS:186, nM:4.
• A mate that runs past the chromosome start maps to nothing in a pair. With filters at 0.3, only the other mate maps (flag 153, 100M, AS:98).
• A single-end read past the chromosome start is unmapped "other".

Change

All in src/align/stitch.rs, plus the removed argument at four calls in src/align/read_align.rs.

• The two extensions at the mate boundary and the two at a transcript's leaf take their flag from ext[iFrag][iEnd]. ext_frag maps a seed's mate_id to that fragment; single-end seeds carry mate_id 2 and use fragment 0.
• An extension that hits a chromosome edge rejects the branch in the stitcher: return false at the boundary, break 'leaf at the leaf.
• finalize_transcript loses its imate parameter and no longer forces any end. Its check for that case moved with it.
• wt_final_score is not floored at 0 any more. An end-to-end extension can take a transcript below zero, and the window's score gate has to see that.

In Local mode every extension flag stays false, as before. Removing the floor does reach Local, for a transcript whose genomic-length penalty takes it below zero; the numbers below show no change there, on these reads.

Measured against STAR

Fields 1-11 plus NH, HI, AS, nM.

Crafted reads (reads with mismatches at the ends, with both mate orders, and reads that cannot reach an end), this branch:

Run	Records STAR / here	SAM and Log.final.out counts
Audit fixture e1.fq/e2.fq	2 / 2	identical, same order
Pairs, EndToEnd	12 / 12	identical, same order
Pairs, EndToEnd, --outSAMunmapped Within	22 / 22	identical, same order
Pairs, Extend5pOfRead1	18 / 18	identical, same order
Pairs, Extend5pOfReads12	16 / 16	identical, same order
Pairs, Extend3pOfRead1	18 / 18	identical, same order
Pairs, Local	20 / 20	identical, same order
Single-end, EndToEnd	18 / 18	identical, same order
Single-end, Extend5pOfRead1	20 / 20	identical, same order
Single-end, Extend3pOfRead1	20 / 20	identical, same order
Single-end, EndToEnd, no mismatch limit	19 / 19	identical, same order
Pairs, EndToEnd, filters 0.3	15 / 18	counts identical; 3 extra unmapped-mate records, which finding 2 removes (15 / 15 with both)

On perf/hotpaths the audit fixture differs, pairs under EndToEnd have 16 records against 12, and Extend3pOfRead1 has 20 against 18.

Fuzz: 4000 repeat-rich pairs, and their first mates for single-end. Reads differing from STAR:

Options	perf/hotpaths	This branch	This branch merged with finding 2
Pairs, EndToEnd	1026	651	0
Pairs, EndToEnd, filters 0.3	2353	2012	0
Pairs, Extend5pOfRead1	662	533	0
Pairs, Extend5pOfReads12	840	566	0
Pairs, Extend3pOfRead1	769	545	0
Pairs, Local	523	523	0
Single-end, EndToEnd	1	0	0
Single-end, Local	0	0	0

So this branch alone does not bring pairs to zero. The differences left are the half-mapped reads of finding 2, which is why Local, which this PR does not touch, stays at 523. The merge's Log.final.out counts are identical to STAR's for every row but single-end EndToEnd, below.

The merge also passes cargo test --release (831 passed).

Not the same as STAR

On the single-end EndToEnd fuzz, the alignments are identical but 7 of the 335 unmapped junk reads carry another label:

Unmapped reason	STAR	Here
too short	117	124
other	218	211

I do not know why. I tried treating a negative score as no alignment: that gave 121 / 214, and three crafted reads scoring 0, -1 and -2 said STAR calls those "too short", so the idea was wrong and I removed it. The crafted fixtures and the audit's reads do not show it. On perf/hotpaths, a read past the chromosome start was "too short" where STAR says "other"; that case is fixed and tested.

Deliberate divergences

None, so no DIVERGENCE.md entry and no sign-off needed.

Tests

tests/end_to_end_pairs.rs, all five failing on the old src/, with expected values from STAR's output on the same fixture:

• inner_ends_are_extended_and_charged
• outer_ends_are_charged_to_the_pair
• a_mate_past_the_chromosome_start_cannot_be_clipped_into_a_pair
• one_sided_modes_extend_only_their_inner_end
• single_end_read_past_the_chromosome_start_has_no_alignment

cargo test --release (824 passed), cargo clippy --all-targets and cargo fmt --check are clean on macOS.

Not checked

• test/parity_suite.sh was not run.
• No real dataset: all numbers are from synthetic reads.
• Speed was not measured. EndToEnd now extends during stitching, on every leaf, rather than once at the end. Local mode takes the same path with the flag off.
• --peOverlapNbasesMin was not run together with --alignEndsType.
• Chimeric detection was not run together with --alignEndsType.

🤖 Generated with Claude Code

[BenjaminDEMAILLE]

Moved to scverse#368, now that scverse#346's branch is copied there.