D12: Single-pass precision/recall alignment with in-alignment FN dropping

[TimD1]

Summary

Replaces vcfdist's greedy, retry-based false-negative (FN) dropping with a single-pass, in-alignment mechanism, then wraps it in a lightweight outer re-align loop to recover swallowed neighbors.

• The truth graph becomes a DAG: each truth variant gains a parallel reference-allele bypass node. Taking the bypass = declaring that variant FN, at cost skip_cost = ceil((1 - ct) * len) (len = variant edit size). One weighted alignment therefore identifies the missed variants directly — no per-candidate re-alignment search.
• The forward pass is now a weighted-wavefront shortest path (Dial's bucket queue); the backtrack labels bypassed variants FN and excludes bypass regions from per-sync-group credit.
• evaluate_variants keeps an outer loop: after a pass, drop the largest detected FN indel/SV and re-align the rest so its swallowed neighbors get clean credit. This replaces the old inner --max-retries exclusion search (which did up to max_retries alignments per iteration just to pick what to drop) with a single detection alignment per iteration.
• Removed: the retry/exclusion machinery, --max-dist/-md, and --max-retries/-mr.

Design and implementation notes: docs/superpowers/specs/2026-07-17-single-pass-prec-recall-design.md, docs/superpowers/plans/2026-07-17-single-pass-prec-recall.md.

Behavior vs. dev (chr20, GIAB, default -ct 0.7)

Not bit-identical to dev by design — FN dropping is now always-on and globally optimal for the bypass cost function. Four-way comparison (TRUTH_TP, recall):

Variant	SNP	INDEL	All TP	vs full-retry
this PR (single-pass + loop)	80160 (0.9704)	8425 (0.6102)	88585	−10
dev full-retry (--max-retries 5)	80160 (0.9704)	8435 (0.6110)	88595	baseline
dev no-drop (--max-retries 0)	80082 (0.9695)	8430 (0.6106)	88512	−83
pure single-pass (no outer loop)	80128 (0.9700)	8426 (0.6103)	88554	−41

• SNPs exactly match full-retry; overall within −10 TP (all INDEL), and +73 over no-dropping.
• The ~10-INDEL residual is a deliberate heuristic trade: this loop drops the largest FN first, whereas full-retry drops the edit-distance-minimizing FN. The remaining entangled-locus credit dilution is the spec §8 residual (characterized, not a bug — exercised by integration scenario E).

Runtime (chr20)

Total wall-clock is a wash — ~1.7s for all variants (dominated by reading ~0.44s + clustering ~0.81s + writing ~0.35s). Isolating the alignment-eval phase: single-pass+loop 0.157s vs full-retry 0.126s vs no-drop 0.114s — i.e. single-pass is marginally slower here, because chr20's small, well-separated variants rarely trigger deep retries and the weighted wavefront costs a bit more per alignment. The expected runtime win only appears where the old loop actually re-aligns many times (SV-dense / large clusters). Validation at that scale (analysis-v3) is still pending a linux-64 environment.

Testing

• Unit (gtest): skip_cost() including large-len regression; existing suite green (2/2).
• Integration (pytest-workflow, tests/integration/): new synthetic-reference scenarios — (a) large FN SV adjacent to SNPs → SNPs TP + SV FN (swallowing recovery), (b/c) partial-match TP at ct=0.7 vs FN+FP at ct=1.0, (d) co-located spurious call → truth FN + query FP with no distortion, (e) entangled-overlap residual (characterization). All validated against the real binary.

Disclosures / follow-ups

• RAM estimate factor changed (src/cluster.cpp): g.max_dist → g.max_size. max_size defaults to 10000 vs old max_dist 1000, so the per-supercluster memory estimate is ~10× larger. Correctness-safe (only makes thread/RAM bucketing more conservative; never drops work), but may reduce parallelism on genome-wide runs with large superclusters. Follow-up: base the factor on an actual wavefront score bound.
• Pre-existing bug (not introduced here): -p <prefix> overwrites input VCFs when the prefix collides with an input file's basename (main.cpp:54-55). Worth a separate issue.
• Broader accuracy validation on analysis-v3 (linux-64) remains outstanding.

Closes #91
Part of #55