feat: after_tool_call → POST /audit (success-only evidence) + config resolution

[olivrg]

Implements the after_tool_call → POST /audit hook (issue #11 Part B) and wires up operator config resolution so the evidence path is reachable in production.

after_tool_call → /audit

• Claims the evaluation_id from the correlation registry (claim miss → no POST; the pending evaluation expires server-side as designed).
• Maps status from event.error (error/success), plus duration_ms and result.
• Success-only, config-driven evidence: per-tool rules with explicit array-segment paths (["url","host"], not dotted) extracted from event.result; a missing path skips the entry.
• Observational/best-effort — never throws across the hook. The payload is hardened so a call that ran always finalizes its audit + counters rather than decaying into a false evaluation_expired:
  • serialization-safe via a single JSON.parse(JSON.stringify(...)) snapshot (no probe-vs-send TOCTOU; tolerates throwing getters / BigInt / circular refs),
  • per-entry evidence filtering (one bad entry doesn't drop the rest),
  • size budgeting under the 1 MiB sideband limit: shed result → trim evidence → truncate (UTF-8-safe) then omit an oversized error, keeping the always-serializable core.
• 200 already_finalized (and any 2xx) is treated as success by the client; we audit unconditionally.

Config resolution

• parseConfig validates api.pluginConfig with zod (defaults mirror openclaw.plugin.json); the evidence config is now consumed at runtime.
• Fail-closed on misconfig: invalid config or a missing adapter token registers blocking before_tool_call + before_install gates and logs loudly — it never skips hook registration, which would leave the tool surface ungoverned. Host-independent (does not rely on the gateway aborting on a register() throw).

Tests

• TDD throughout. after-tool-call.test.ts (claim/status/evidence/serialization/budget/error-truncation), config.test.ts (zod validation + defaults), and an end-to-end index.test.ts proving configured evidence rules reach the /audit body through real plugin registration.
• pnpm verify green: 162 tests (vitest runtime + typecheck projects), typecheck clean, build success.

Not in this PR

before_install → /install-scan remains a stub (next task); Part C (live e2e + npm publish) follows.