[cjfields]

Ports Marco-Sola-Lab/WFA2-lib#140 (cda19a0, already in the pinned C commit). I had listed it as "speed only, n/a", but Rust had the same problem, and worse. Relevant to #31 and to dada2-rs, whose per-thread aligner sees mixed read lengths in high memory.

Problem

In high memory (CIGAR), init_alignment calls WavefrontSlab::reap_cigar. It walked every wavefront the slab had ever created, including deallocated ones. A long alignment leaves thousands of them, so every short alignment that followed paid for the whole walk. sample put about 70% of a mixed run in reap_cigar.

Fix

• The slab records the indices it hands out (busy).
• When there are no mixed lengths (mixed_lengths, C's dirty), reap_cigar returns only those indices. That's safe because every other live wavefront is already on the free list.
• Recording costs about 2% on uniform short workloads, so it's done only while the slab has at least 4096 wavefronts, which happens only after long or divergent alignments. A busy set of a quarter of the slab or more still takes the ordered full walk.
• A debug_assert covers the invariant: a deallocation implies mixed lengths.

Results

M2 Pro, median of 7, alignment time in ms, on C's benchmark data (10 × 10 kb then 200k × 100 bp, 5% error, gap-affine):

Data	Mode	C before #140	C after #140	Rust main	This PR
mixed	high	700	421	1231	365 (−70%)
mixed	med	774	782	651	653
short only	high	362	376	322	318
short only	med	609	619	565	556

Interleaved workload (each 10 kb pair followed by 300 × 100 bp), median of 5:

• dada2-like (-5,4,0,8, ends-free 0.1, banded ±16, corrected): 735 → 209 ms;
• plain gap-affine: 56 → 45 ms.

Validation

• Output: byte-identical to main on the interleaved set for every metric, in high/med/ultralow and score-only, plus corrected ends-free (affine and linear, banded) and X-drop.
• Differential suite:
  • C mode: 1262 cases, only the 20 known Known issue (C bug): Z-drop in med/low backtraces from uninitialized cells (high vs med differ) #22 differences;
  • corrected mode: identical to the run before this change.
• Tests: new unit test test_reap_cigar_busy_after_large_slab. Tests pass in release and debug, so the debug_assert ran; clippy is clean.

PERFORMANCE_LOG and PROGRESS are updated; C #140 is now marked as ported.

🤖 Generated with Claude Code