[cjfields]

Follow-up to #40 (part of #31 / #32). Makes Compat::Corrected ends-free with match < 0 nearly as fast as C mode, so dada2-rs (HPCBio/dada2-rs#51) can re-measure it against NW.

Changes

• Faster extend kernel. The corrected-mode overhead (+26–28% instructions, +26–51% time over C mode) wasn't from the extra steps: those were only +0.8% on long.seq. It came from the _best extend kernel, which called endsfree_left (two Options, merged) on every diagonal. Samply showed extend at +74%. The kernel now walks the offsets like the C-mode kernel and runs the cheap termination_endsfree check first, counting the bases left only at an end.

• Exact pruning. From diagonal k at score s, any completion totals at least s + M·|k − k_end| (k_end = |T| − |P|). The bases left differ by k − k_end all along a diagonal, and each of them costs at least M. Once an end candidate exists, every component at each score is trimmed to the diagonals that could still beat it (ties kept), and the dropped cells are set to OFFSET_NULL, so later steps can't read them (Wavefront::trim_nulled). This runs only with no heuristic or BandedStatic, because the other heuristics read the wavefronts.

• Bug fix (FIXED_BUGS 14c). Null steps no longer end a corrected search while a candidate is pending. With BandedStatic, the band (or the pruning) can empty every wavefront before begin-free seeds outside it are reached, and the search then returned a worse end. Example: SW 1 instead of 2 on tests/data/corrected_banded_seeds.seq, band -10,200.

• Fixes Debug panic (subtract overflow) in backtrace when both sequences are empty #41: a fresh aligner given an empty pattern and empty text panicked in debug builds (usize underflow of the backward CIGAR write offset; release builds wrapped and happened to be right). cigar_buffer_len adds one slot to the C-sized 2·(|P|+|T|) buffer at all three sizing sites (FIXED_BUGS 15). test_empty_sequences_cigar covers edit/affine, high/med/low, end-to-end/ends-free and BiWFA; it fails without the fix.

Results

Instructions retired on long.seq (300 × 2 kb), CIGAR / score-only. I used instruction counts because the machine was loaded for part of this. Wall time is best of 5, interleaved, at load ~3.

Config	main C	main corr.	branch C	branch corr.
affine -5,4,0,8, ends-free 0.1 ×4	14.54 / 14.25 G	18.30 / 18.01 G	14.54 / 14.25 G	14.87 / 14.58 G
same, ms	1092 / 715	1451 / 1054	1103 / 717	1181 / 796
ends-free 1 ×4, ms	1713 / 1230	2344 / 1856	1710 / 1237	1854 / 1382
extension, ms	145 / 119	198 / 169	145 / 120	159 / 130
affine2p -1,6,2,4,12,1, ends-free 0.1 ×4, ms	1444 / 1008	1814 / 1372	1451 / 1008	1533 / 1087
ends-free 0.1 ×4, banded-static ±200, ms	904 / 560	1160 / 816	880 / 561	956 / 613
global, ms	128 / 105	127 / 106	128 / 105	129 / 105

• The corrected-mode overhead over C mode is now +2–2.5% instructions (+7–12% time), down from +26–28% (+26–51%). C mode is unchanged.
• Pruning on its own is modest: 0.3–2%. On long.seq the first end found is usually near-optimal already. On pairs with random flanks and whole-length free ends, corrected mode is still +15–55% (up to +160% on short pairs) over C mode. In those cases C stops at a wrong, Known issue (C parity): ends-free with whole-length free ends and match < 0 returns the no-overlap alignment #38-style first hit, so the extra search is what makes the result optimal.

Verification

• Pruned vs unpruned output: byte-identical (both builds include the 14c fix):
  • 360 configs over 506 generated pairs (#37/#38 cases included);
  • 240 configs over long.seq;
  • configs cover linear/affine/affine2p, five ends-free/extension spans, high/med/low, CIGAR and score-only, and no heuristic plus two banded-static bands.
• New tests:
  • random_long_against_oracle: 300 pairs up to about 200 bp with flanks, checked against the DP oracle in all memory modes. It fails if the pruning band is one diagonal too narrow.
  • banded_static_pending_seeds: fails without the 14c fix.
  • A trim_nulled unit test.
• cargo test --release and clippy are clean. The new code is rustfmt-formatted.
• With the Debug panic (subtract overflow) in backtrace when both sequences are empty #41 fix, the full suite also passes in a debug build (on main, three corrected_endsfree tests hit Debug panic (subtract overflow) in backtrace when both sequences are empty #41 in debug). wfa_utest.sh 60/60.

Fixes #41

🤖 Generated with Claude Code