Phase 3 — Edge cases: tighten mode + gapped fallback parity + real-data validation

[k1sauce]

Part of #1. Depends on #4.

Scope

After Phase 2 brings chr22 MAPQ to 97%+, close the remaining gap by porting BT2's score-tightening and verifying the gapped-fallback path matches BT2's behavior in default mode.

Deliverables

Score tightening

BT2 default mode does NOT use tightening (`tighten > 0` is gated on `-M` mode which is off by default). But there are subtle score interactions:

• Audit BT2's `extendSeedsPaired` for any score-threshold updates that fire in default mode (not just M-mode).
• Specifically check `ominsc_cur` recomputation at `aligner_sw_driver.cpp:2173-2205` and confirm whether it ever fires when `tighten == 0`.

Gapped fallback parity

• Our `paired_descent.rs` has a Phase-2 gapped fallback that fires when ungapped finds nothing for both mates. Verify BT2's gapped path fires under the same condition.
• Verify gapped extension score formula matches BT2's exactly.

Real-data validation

• Run `scripts/real_data_validate.sh` post-Phase-2.
• Compare to current 96.8% baseline.
• If real-data MAPQ <99%, investigate (different read characteristics: longer reads, quality variation, indels).

Exit criterion

• chr22 MAPQ → 99%+ at hi-cap.
• Real-data MAPQ → 99%+, or documented why it can't reach.

Risks

• Real-data has indel-heavy reads that exercise gapped extension more than the synthetic. The synthetic-fitted descent may not transfer cleanly.
• BT2's actual default mode might have undocumented behaviors we discover here.