[cjfields]

Part of #31, prompted by HPCBio/dada2-rs#51's finding that begin-free gaps make each ends-free alignment about 3× slower. Details are in the #31 comment and PERFORMANCE_LOG.

Findings

• The lead is confirmed: begin-free seeded diagonals stay in the wavefront. At 240 bp, 2%: 35 → 604 cells and 10 → 43 non-null steps per alignment.
• Pruning them early doesn't help. I tried an initial upper bound for corrected mode's exact pruning: output was identical, but time was 2–18% worse. With a match reward, a cell's lower bound grows only |match| per diagonal from the end diagonal, against 18 per difference, so the seeded diagonals can't be ruled out until the search is nearly done. Not included here.
• The time follows the step count: about 170 ns per extra non-null step, against under 1 ns per cell of kernel work. The per-step costs are spread out; the two clearest are fixed here.

Changes (separate commits)

1. 494602c: init_ends fills of up to 8 offsets use volatile stores instead of slice::fill. fill lowered to a memset_pattern16 call per input per step, whose setup cost far more than the one or two stores.
2. 9034dc5: apply_heuristic_cutoff returns at once for banded-static when the wavefront is inside the band. Compute already clamps to the band; only begin-free seeding reaches past it. That path doesn't use steps_wait, so skipping it changes nothing.
3. 704a5e4: PERFORMANCE_LOG and PROGRESS notes.

Results

M2, median of 5, in ms, main (bef359a) → this PR. Output identical in every case.

data	linear, begin off	linear, begin on	affine, begin off	affine, begin on
240 bp, 2%	5.6 → 4.8	16.4 → 14.5	8.0 → 7.0	25.1 → 22.6
240 bp, 10%	16.6 → 14.1	47.6 → 41.1	23.6 → 20.2	72.2 → 60.7
1450 bp, 2%	58.8 → 52.9	190.3 → 172.8	79.6 → 70.3	257.5 → 226.8
1450 bp, 10%	164.4 → 149.5	479.1 → 437.5	220.0 → 193.9	659.5 → 586.0

This is 9–16% faster on dada2-like alignments. The utest and long gap-affine controls are at 0.99× and 0.96×. The 3× from begin-free remains: it's the step count of an exact WFA with seeded diagonals.

Validation

• 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.
• Differential suite, corrected mode: unchanged from the previous run.
• Tests: pass in release and debug; clippy is clean.

🤖 Generated with Claude Code, with edits by @cjfields