… + actor_id

Closes the KR-MCP-STOP-CONTROL arc. Three deliverables in one PR:

  D1. actor_id field on Caller dataclass + mcp_callers.yaml loader
      - Optional uuid string. Backwards-compatible — existing
        callers (no actor_id key) load with actor_id=None.
      - Loader rejects malformed UUIDs (fail-CLOSED skip + WARN).
      - Accepts unhyphenated form; normalizes to canonical form.

  D2. kora_cli/clients/kora_control_writer.py — Python wrapper
      around public.issue_kora_control SECDEF (substrate
      packages/db/migrations/0090_kora_control_secdefs.sql,
      shipped substrate-side at commit 2abf095, May 21 2026).
      - Raw asyncpg through the IsoKron pool, mirroring
        KoraControlReader's pattern. Bucket-spec drift: spec
        said "IsoKronMCPClient.invoke('substrate__issue_kora_
        control', ...)" but no MCP wrapper exists in
        sea-mcp-server/src/tools/; the canonical access path
        for this SECDEF is direct asyncpg.
      - dry_run mode short-circuits before substrate; predicted
        shape uses zero-UUID placeholders + sequence=-1.
      - 10s client-side asyncio.wait_for ceiling
        (ISSUE_KORA_CONTROL_TIMEOUT_SECONDS) per §4 Q3.
      - Fail-CLOSED on: missing actor_id, invalid (level, kind),
        missing pool, asyncpg PostgresError (→ SubstrateRejected
        with sqlstate preserved), client-timeout (→ Writer
        Timeout).
      - Module-level current_kora_control_writer() accessor
        bound to the active IsoKronMemoryProvider singleton.

  D3. kora__request_stop MCP tool — L1/L2 only.
      - JSON schema constrains level to [1, 2]; executor ALSO
        refuses L3-L5 with explicit operator-on-machine message.
      - confirm_token must equal daemon's daemon_session_id
        (newly added on DaemonCoordinator at construction +
        surfaced via get_status()). Mismatch → -32602.
      - Caller cap-gate on "kora__request_stop". DISTINCT from
        kora__request_pause (verified by test) — operator can
        grant pause/resume without granting stop.
      - actor_id required: -32001 actor_id_required_for_stop
        (new error class _ST2_ActorIdRequired in mcp_tools.py
        with dedicated handler in mcp.py).
      - Audit emit: {level, kind, dry_run, actor_id_present} —
        reason text NEVER in audit details per §4 Q4 ruling.
      - Returns immediately after substrate write succeeds;
        enforcement is async via Kora's poll loop.

Additional surface:
  - DaemonCoordinator._daemon_session_id (hex uuid4, per-process
    stable) — exposed via .daemon_session_id property + the
    get_status() JSON key. Read by kora__daemon_status for the
    confirm_token mint flow.
  - _jsonrpc_error extended to accept optional `data` field for
    structured error envelopes (used by the actor_id_required
    branch to surface caller_actor_kind + remediation text).

§4 PM-opens resolved:
  Q1 — SECDEF verified at packages/db/migrations/0090_kora_
       control_secdefs.sql; full signature pasted into
       kora_control_writer.py module docstring §1.
  Q2 — operator-pastes actor_id from substrate actor_registry
       (must be substrate-resolvable per SECDEF body checks).
  Q3 — 10s wait_for ceiling, SECDEF's 60s statement_timeout is
       the substrate-side backstop.
  Q4 — {level, kind, dry_run, actor_id_present, result_tag}
       audit keys. Reason text omitted.

Tests:
  - tests/kora_cli/clients/test_kora_control_writer.py: 14
    writer-level unit tests (dry_run, missing actor_id,
    invalid level/kind, no pool, SubstrateRejected wrap,
    timeout wrap, current_writer accessor states).
  - tests/kora_cli/test_listeners/test_mcp_tools_stop.py: 16
    request_stop tool tests (L1/L2 success, L3-L5 refused,
    L0 refused, confirm_token mismatch, empty token,
    no capability, no actor_id, dry_run, substrate-error,
    no coordinator, no provider, audit-omits-reason, SECURITY
    bearer-never-in-envelope).
  - tests/kora_cli/test_listeners/test_mcp_caller_auth.py: 9
    new tests for actor_id parsing (default-None, valid UUID,
    malformed UUID, non-string, unhyphenated, mixed entries).

Full regression: 9051 passed (matches base+57 new). 47 failed
identical to ST1 baseline = zero new regressions. All failures
are pre-existing in tests/agent/test_anthropic_adapter.py +
tests/kora_cli/test_web_server*.py — unrelated to this PR.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>