feat: `vouch metrics` — first-class observability

[plind-junior]

Labels: enhancement

What you're trying to do

vouch status shows artifact counts — useful for "is the KB alive?" but not "is the review gate working?". Operators want time-series-able signals: how many proposals get approved, how stale the corpus is, how long claims sit pending, who is proposing/approving.

ROADMAP 0.3 calls for observability hooks. All the data already lives in audit.log.jsonl + the artifact files — no new on-disk state needed.

Suggested shape

vouch metrics                # human table
vouch metrics --json         # machine-readable (Prometheus textfile compatible)
vouch metrics --since 30d    # window the audit log

Emits:

• approval_rate — approve / (approve + reject) over the window
• citation_coverage — fraction of approved claims whose evidence ids still resolve
• stale_ratio — fraction of claims past lint's stale threshold
• proposal_lag_p50 / p90 — time from proposal.*.create → .approve
• actors — top N by proposals/approvals

Acceptance

• vouch metrics --json produces a stable schema documented in docs/metrics.md.
• Counts derive purely from audit.log.jsonl + artifact files (no new disk state).
• Test asserts metrics against a fixture KB with known stats.

Out of scope

• Pushing metrics to Prometheus/Datadog (do that with a sidecar reading the JSON output).
• Long-term retention / TSDB.

---