STAR-compat batch (--alignEndsType, --outSAMorder, chimeric/transcriptome) + PGO CI fix

[Psy-Fer]

STAR-compat batch (--alignEndsType, --outSAMorder, chimeric/transcriptome) + PGO CI fix

Works through five STAR-faithfulness items surfaced in an older deep-dive review, plus fixes the flaky linux-x86_64-v4 release build. Two logically independent changesets on one branch — split into two commits if you'd rather land the CI fix on its own.

Gate (whole branch): 560 lib + 26 integration = 586 tests pass · clippy --all-targets 0 warnings · fmt clean. Yeast SE Local 8788/8926 and PE 8390 both-mapped / 0 half-mapped — identical to baseline.

---

Part 1 — STAR-compatibility batch

1. --alignEndsType (read-end extension mode)

Implements Local (default), EndToEnd, Extend5pOfRead1, Extend5pOfReads12, Extend3pOfRead1 — STAR's alignEndsType.ext[iMate][iEnd] boolean matrix (Parameters.cpp, extendAlign.cpp, stitchWindowAligns.cpp).

• extend_alignment gains an extend_to_end branch that forces extension over the full remaining read (no max-score tracking, no mismatch-limit break; chromosome-boundary hit kills the transcript), matching STAR's extendToEnd path.
• finalize_transcript derives per-end flags from ext[imate][iEnd]. Because it runs per-mate-slice, original_is_reverse equals STAR's (Str != iMate), so the mapping reduces to ext[imate][is_reverse] (start) / ext[imate][!is_reverse] (end). Internal stitch extensions stay Local; forcing is applied only at the read-end extension (additive scoring makes deferring to finalize equivalent to STAR's in-stitch forcing).

Validation: Local default byte-identical to baseline (SE 8788/8926, PE 8390/0). --alignEndsType EndToEnd vs STAR --alignEndsType EndToEnd on 10k yeast: 98.4% position agreement (7924/8054), rustar 8821 vs STAR 8817 mapped — at parity with Local's tie character. +2 unit tests (test_extend_to_end_*).

2. --outSAMorder

--outSAMorder (Paired | PairedKeepInputOrder). rustar's align pipeline already emits records in exact input (FASTQ) order regardless of thread count (verified empirically on a 4-thread run — identical ordering). The flag is now accepted (both values, effectively a no-op) for STAR/pipeline compatibility; unknown values are rejected. +3 param tests.

3. PE chimeric diffMates gap relaxation

The specific "gap check wrongly filters inter-mate chimeras" claim doesn't hold under rustar's architecture: detect_chimeric_old runs strictly per-mate, and inter-mate is handled by detect_inter_mate_chimeric, which imposes no inter-mate read-gap. Still made the function STAR-faithful: detect_chimeric_old_impl gains an optional mate_boundary and waives the read-gap across it (STAR ReadAlign_chimericDetectionOld.cpp:67-71). The public detect_chimeric_old and all per-mate callers pass None — zero behavior change. +1 unit test asserting the gap is enforced without a boundary and waived with one.

4. Chimeric score-drop read_len (verify)

Verified self-consistent, no fix needed: the score-drop gate uses the length of the read passed in, so it scales correctly for both modes — per-mate length for rustar's PE pools (an intra-mate chimera spans one mate) and combined length for a combined read (matching STAR's readLength[0]+readLength[1]). Added a clarifying comment.

5. Multi-exon minus-strand transcriptome projection (test gap)

Added the missing regression test for a 2-exon reverse-strand transcript projection (align_to_one_transcript tr_strand==2 branch): exon-order reversal, t-space/read-coord inversion relative to transcript/read length, and N-op stripping. Passes first try — the projection code was already correct.

---

Part 2 — CI fix: flaky linux-x86_64-v4 (and preventive neoverse-v1)

Symptom: the linux-x86_64-v4 release job intermittently died with Illegal instruction (core dumped) / exit 132 in the PGO training run; re-running sometimes passed.

Cause: the PGO instrumented binary was built at the shipped ISA (-Ctarget-cpu=x86-64-v4, AVX-512) and then executed to collect the profile. GitHub's ubuntu-latest runners don't reliably have AVX-512, so on a non-AVX-512 runner the training run hit an AVX-512 instruction and SIGILL'd — flaky by which runner the job landed on.

Fix: decouple the training ISA from the shipped ISA. scripts/pgo-build.sh gains PGO_TRAIN_EXTRA_RUSTFLAGS (defaults to PGO_EXTRA_RUSTFLAGS), applied to the instrumented build only. The release matrix trains at a runner-safe ISA and ships at the optimized one:

target	train ISA	ship ISA
linux-x86_64-v4	x86-64-v3 (AVX2)	x86-64-v4 (AVX-512)
linux-aarch64-neoverse-v1	generic (armv8-a)	neoverse-v1 (SVE)

PGO profiles are behavioral (block/branch frequencies) and transfer across -Ctarget-cpu levels, so the shipped binaries remain fully optimized. The neoverse-v1 entry gets the same split preventively — identical failure mode on any non-SVE aarch64 runner.

Validation: ran a full local train=v3 / ship=v4 PGO cycle — completes, ships a working v4-optimized binary (the -pgo-warn-missing-function notices are the expected benign cross-ISA function mismatches). rustc confirmed to accept -Ctarget-cpu=generic for aarch64; workflow re-validated as valid YAML.

---

Files

• src/params/mod.rs — AlignEndsType type + --alignEndsType/--outSAMorder params, validation, tests
• src/align/score.rs — align_ends_type on AlignmentScorer
• src/align/stitch.rs — extend_to_end branch, finalize_transcript ext-flag wiring, tests
• src/align/read_align.rs — per-mate imate args to finalize_transcript
• src/chimeric/detect.rs — detect_chimeric_old_impl + diffMates, score-drop comment, test
• src/quant/transcriptome.rs — multi-exon minus-strand projection test
• scripts/pgo-build.sh, .github/workflows/release.yml — PGO train/ship ISA split
• README.md — --alignEndsType and input-order/--outSAMorder feature bullets