[Tracking] BT2 full-descent port: path to 100% MAPQ parity

[k1sauce]

Goal

Achieve 100% MAPQ agreement with BT2 on the chr22 synthetic validation corpus. Stretch goal: same on real-data (NA12878 mitochondrial).

Current state

• Synthetic chr22 MAPQ: 93.9% (baseline) / 94.2% (--seed-hit-cap 1000 -D 1000) / 94.1% (--joint-descent).
• Real-data MAPQ: 96.8%.
• Two BT2 precision bugs already fixed: score_min C-style truncation, mapq bin thresholds f32→f64 cast.
• Joint-descent infrastructure in place behind --joint-descent flag.

Root-cause analysis (Phase 0, #2 — closed ✓)

The remaining ~6% gap is BT2's default-mode paired-descent traversal-order short-circuit: `bestUnchosenCScore` comes from the pair BT2 happens to emit during a tightly truncated descent (mean pool size ~1.2 entries), not the cartesian-second-best. Our pool keeps 30+ entries with a different specific secbest. Capping our pool to size 2 doesn't fix it — top-2-by-score picks the wrong entry.

Full writeup in `rusttie.md` (Phase 0 section).

Plan

• Phase 0 — Spike: identify exact BT2 pool divergence root cause #2 — Phase 0 — Spike: identify exact divergence ✓
• Phase 1 — Port BT2's GroupWalk SA-range traversal #3 — Phase 1 — Port `GroupWalk` SA-range traversal
• Phase 2 — Port RedundantAlns + paired anchor iteration #4 — Phase 2 — Port `RedundantAlns` + paired anchor iteration
• Phase 3 — Edge cases: tighten mode + gapped fallback parity + real-data validation #5 — Phase 3 — Edge cases (tighten mode, gapped fallback, real-data)
• Phase 4 — Cutover: make joint-descent the default #6 — Phase 4 — Cutover: make joint-descent default

Checkpoints

• After Phase 0: Smoking gun found ✓ → continue with Phase 1.
• After Phase 1: Single-end metrics unchanged → continue or debug.
• After Phase 2: MAPQ moves to 97%+ → continue or accept new ceiling.

Honest unknowns

1. Phase 0 might fail → resolved, root cause confirmed.
2. The 6% gap might be a combination, not GroupWalk alone — each phase might only close a fraction.
3. BT2 is GPL-3; this must be clean-room reimplementation, not copy.
4. Real-data might not converge even if chr22 does — different read characteristics.