07 — Rollout + per-runner observability (canary via wizard-runner)

Epic: #520 · Behavior change: first user-visible · Depends on: #521–#526

Summary

Add per-pair guardrail telemetry and start a small canary using the existing wizard-runner flag.

Scope

• Tag run telemetry with the resolved router + runner + model: success rate, wall-clock, tool-error rate,
  token usage, YARA blocks. (Cuts by runner and by model both fall out.)
• Structured run logging to file (the TUI owns the console): assistant turns + tool I/O.
• A dashboard comparing runners (anthropic vs pi) and routers.
• Start a small canary by ramping the existing wizard-runner toward pi for a targeted slice, gated on
  metrics. The one-flip revert is wizard-runner=anthropic (today's default) — no new kill-switch flag.

Acceptance criteria

• Telemetry carries the resolved router + runner + model; dashboard live.
• Canary active via wizard-runner; revert to anthropic verified.
• Guardrails at/above the anthropic baseline before any ramp.

Files

• src/lib/agent/runner/*, analytics

Start by reading the existing runner code under src/lib/agent/runner/*, the analytics area, and the current wizard-runner flag behavior. Define the telemetry and structured logging points, then verify the dashboard compares runners and routers, the canary ramps toward pi, and reverting wizard-runner=anthropic works with guardrails at the anthropic baseline.