chore(appkit): production hardening — shutdown re-entrancy + SSE idle keep-alive

* SIGTERM/SIGINT re-entrancy guard in ServerPlugin._gracefulShutdown.
  In interactive dev the OS sends SIGINT and the supervisor follows up
  with SIGTERM, racing two _shutdownCoreServices invocations — observed
  failure modes are Lakebase pool double-close, OTLP batcher double-
  flush, and "process.exit called twice". Guard the second signal with
  a flag and log it at debug.

* Wall-clock idle keep-alive in the executeTask SSE bridge (25 s,
  IDLE_KEEPALIVE_INTERVAL_MS). The engine emits heartbeat events only
  while the executor is actively running; a task waiting on a slow
  downstream call can stay quiet long enough for an AWS/GCP/Cloudflare
  load balancer to drop the idle socket (typical 60 s timeout).
  Belt-and-braces wall-clock interval on top of the engine path,
  unref()'d so it doesn't keep the process alive on its own, and
  cleared from the bridge's outer finally so any error or early-exit
  cleanup path doesn't leak the timer.

* CLAUDE.md: new "TaskFlow Core Service" section documenting the
  mental model, public API on this.taskflow, the executeTask shortcut,
  the when-to-use-which-execution-method table, the recovery pattern,
  and OBO/autoRecover guardrails. Adds vendored @databricks/taskflow
  to Key Dependencies (sha256-pinned via VENDOR.json) and a sixth
  "durable by default for long ops" design principle. Updates Graceful
  Shutdown bullets to reflect the new re-entrancy behaviour and the
  bridge drain on shutdown.

* .claude/references/taskflow.md (new, 404 lines): canonical
  NEVER/MUST/SHOULD reference for TaskFlow usage from plugins. Covers
  when-to-use, task registration, idempotency keys, handler signature
  and ctx.emit, recovery patterns (agent loop, staged pipeline, saga),
  OBO/asUser interaction, shutdown semantics, and conflict semantics
  on duplicate submits. CLAUDE.md links to it as the authoritative
  source for the rules its TaskFlow Core Service section summarises.

Tests:
- src/plugins/server/tests/server.test.ts: re-entrancy guard test
  fires two _gracefulShutdown calls in parallel and asserts
  shutdownCoreServices / abortActiveOperations / server.close each
  run exactly once.
- src/taskflow/tests/execute-task.test.ts (new): two focused tests
  for the wall-clock keep-alive — one asserts the comment frame fires
  per IDLE_KEEPALIVE_INTERVAL_MS window while the engine is silent,
  the other asserts the interval is cleared once the bridge exits
  cleanly so a later advanceTimersByTime adds no new frames.

Deliberately out of scope:
- Register-time OBO+autoRecover hard-error in TaskflowService.task():
  OBO-ness is a property of the *caller* of executeTask (the active
  UserContext scope), not the registration, so a register-time check
  is structurally not available without adding a new isOboOnly flag
  to TaskDefinition. The runtime first-call warning from PR 4
  (oboAutoRecoverWarned set, deduped per (plugin, task)) covers the
  misconfiguration at the boundary where it materialises.
- VENDOR.json source-commit pin: already shipped in PR 3.
- Internal review notes (taskflow-review-findings.md,
  taskflow-review-plan.md) intentionally not committed.

Validation: pnpm -r typecheck ✓, pnpm build ✓, pnpm exec biome check
(touched files) ✓, pnpm exec knip ✓, pnpm test ✓ (126 files / 2307
tests; +1 file / +3 tests vs PR 6).

Signed-off-by: ditadi <victordperd@gmail.com>

Author: ditadi