[cjfields]

Refs #22, #55, #21. These stay open until upstream C resolves them. All changes are behind Compat::Corrected (#32); Compat::C is unchanged. There are four atomic commits:

1. 210f60e: Z-drop ends backtraced at the best cell's score (#22)

On a Z-drop, C backtraces the best cell's diagonal and offset starting at the current score, where that cell needn't exist.

• High memory: the backtrace finds no source and pads with unchecked matches. With default penalties that's usually max-trimmed to empty. With match < 0 the CIGARs are invalid, e.g. 2I58M on a 2 kb, 8%-error pair, and up to 300/300 invalid on long.seq with affine -3,5,3,2 and Z-drop 100,5. Known issue (C bug): Z-drop in med/low backtraces from uninitialized cells (high vs med differ) #22 only described the med side, so this part is new.
• med/low: the backtrace reads a dead cell that C never initialized (Known issue (C bug): Z-drop in med/low backtraces from uninitialized cells (high vs med differ) #22).

Corrected mode backtraces the cell at the score it was found at:

• high memory from the stored wavefronts;
• med/low from the cell's (pcigar, prev), saved when it became the best and pinned through backtrace-buffer compaction (compaction now takes a slice of pinned chains).

2. b7b7889: X-drop/Z-drop score with match == 0 (#55, new)

C's Smith-Waterman score for these heuristics uses swg_match = -1 when match == 0, which is the default for every gap metric. This regression from C's 51fb1aa flipped the old "match-score=1" stand-in, so progress lowers the score:

• X-drop prunes the diagonals that progress most, abandoning nearly every divergent pair (long.seq: typically 0/300 optimal).
• Z-drop's best cell is the start of the alignment.

Corrected mode uses +1 (heuristic::swg_match). With default penalties, X-drop is then optimal on 300/300 long.seq pairs at 50,10 / 100,5 for affine, linear and affine2p, and Z-drop gives real partial alignments.

The #21 tests move to wfa.utest.seq pair 305 with match-reward penalties and X-drop 20,2, so they still check that corrected scores equal C's. Their CIGAR check now uses the SW-score bound, which is meaningful with a match reward.

3. 6374f4d: pass Compat to BiWFA's sub-aligners (Rust bug from #54)

BiWFA's forward, reverse and base sub-aligners were created with Compat::C and only given the parent's heuristic. So the #21 fix (and now #55) never applied under BiWFA; #54's "BiWFA identical to C" was this, not parity. Now:

• BiWFA X-drop has no invalid CIGARs in corrected mode (C: up to 213/300).
• With default penalties, BiWFA X-drop is optimal on 168–303 pairs (C: 0–240).

4. 70e6f2d: BiWFA Z-drop ends at the best cell (#22)

BiWFA's base step backtraces a Z-drop in the same way as C, and also starts in the sub-problem's end component (I/D at a breakpoint), though the best cell is an M cell. Corrected mode starts in M at the best cell's score. Across BiWFA Z-drop configurations:

• no invalid CIGARs (C: up to 64);
• with default penalties, far fewer empty results (long.seq 50,10: 0, C: 300).

Results

• 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: every difference is X-drop, Z-drop, or ends-free/extension with match < 0 (the existing fix). Banded, adaptive and core cases are identical to C.
• Z-drop 20,1 / 50,1 / 50,10 / 100,5 × default and -3,5,3,2 × global / extension / ends-free × wfa.utest.seq and long.seq:
  • every partial CIGAR is valid;
  • high = med except co-optimal ties (an empty CIGAR vs 1M/2M, both scoring 0).

Tests

• tests/corrected_zdrop.rs:
  • C mode's invalid CIGAR is reproduced; corrected high/med/low agree;
  • a compaction case that fails without the pin (checked by mutation);
  • BiWFA.
• tests/corrected_xdrop.rs: now also checks BiWFA.
• test_compute_sw_scores_c_formula: covers both stand-ins.
• cargo test passes in release and debug; clippy is clean.

Docs: the Compat list, the XDrop/ZDrop rustdoc, the README, and FIXED_BUGS 14e, 14f and 14g.

🤖 Generated with Claude Code