CLI:
- Drop `external-agents refresh` (pulled agents.yaml from GitHub main).
  The bundled registry ships via npm — remote-override was novel but
  invited drift and the source of truth is now the versioned package.
- Add `external-agents audit [--provider P] [--json]`. Force API
  round-trip against every entry with a generate_new transport, classify
  the response (healthy / needs_auth / model_unavailable / rate_limited)
  and write the verdict to state.json. Runs concurrent per-provider so
  a slow bucket does not block the fast ones. First real use case:
  detects Cerebras entries pointing at models the account can't reach.

UI:
- Replace the noisy "Speed (tok/s)" column with "Success" — shows
  `<ok-count> / <fail-count> fail` from stats.outcomes for the last 24h.
  Color-coded: all-ok green, mixed orange, all-fail red, dim zero.
  Speed was misleading — duration_ms includes network + prompt processing,
  not pure generation time. Success/fail is the actionable signal.

README: replace "Keeping the list fresh" (refresh-based) with
"Keeping the list honest" (audit-based) — better framing of what the
command actually does.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>