Adaptive routing memory: learn task-class priors from outcomes

[anandgupta42]

Parent: #1134

Learn from mistakes — adaptive routing memory:

• Persist per-task-class outcomes locally (class → primary attempts, gate passes, escalations). Task class = coarse heuristic (e.g. task-type keywords, project size bucket), not prompt text.
• Routing consult: classes whose primary success rate is below a threshold (default e.g. <20% over ≥5 samples) skip the primary and go straight to escalation; surfaced in the session UX ("skipping primary: this task class escalated 9/10 recent runs").
• User control: view (route report includes the memory), reset, and disable (route.adaptive: false).
• Keep it transparent and simple: counts and thresholds, no opaque scoring in v1.
• Tests: threshold behavior, decay/window, reset, disable.