v2.58.0 — autopilot Tier-2 (Wave G)

Tier-2: quality & trust. Makes the operated autopilot pipeline not just run but stay honest, rigorous, and reactive — all buildable without external IdP/sandbox creds.

• SLA auto-escalation — the deadline clock now acts: slaState() (ok / at-risk / breached) + autoEscalateStale() escalate breached runs once each (idempotent, audited sla-escalated); new board cron sla.escalate (5 min) sweeps every tenant + fires email/push.
• Calibrated confidence + closed loop — calibration() reliability curve + ECE (is the AI's confidence honest?), suggestFloor() proposes a data-driven confidence floor; new admin endpoint GET /api/autopilot/calibration.
• Eval rigor — ×3 median CI gate — vertical-scorecard.mjs --median N --ci THRESHOLD takes the median of N passes (kills single-run noise) and exits non-zero below the bar. A real regression gate.
• Sequential review pipelines — stageProgress() surfaces multi-gate runs as intake→QC→review→submit, attached to /api/autopilot/run as stages.
• Flake fix — Wave F retention test boundary race resolved; lib suite 299/299 green ×5.

Implements docs/plans/PLAN-autopilot-tier2.md. 6 new lib tests, no functional regression.