App lifecycle: async open/close + DI seam

[sagi5060]

Closes #1

What

• App.open(): an async classmethod context manager that builds the app,
  runs load(), starts MCPLifecycle, and guarantees aclose() runs on
  exit (even on error).
• App.aclose(): closes the SessionFactory's Redis client and shuts down
  MCPLifecycle. Idempotent — a second call is a no-op, guarded by a
  _closed flag.
• App(session_factory=...): the constructor now accepts an optional
  SessionFactory override so tests can inject a fake (e.g. one wrapping
  fakeredis) instead of building a real Redis client from settings.
• serve.py: create_app() now wires App.open()/aclose() through a
  FastAPI lifespan, so compose stop (SIGTERM) closes the Redis client and
  MCP servers cleanly instead of leaking them.

Why

App had no shutdown path — the Redis client and MCP lifecycle were never
closed, and there was no seam to swap infrastructure in tests. This lands the
minimal DI + lifecycle surface the issue asks for, without adding any new
catalog/execution logic (agentdeck still just owns configuration).

Decisions / notes

• MCPLifecycle is a process-wide classmethod registry (not per-App), so
  open()/aclose() call its startup()/shutdown() directly — both are
  already individually idempotent, matching App.aclose()'s idempotency.
• serve.py's lifespan calls deck.load() a second time (after open()
  already ran it once) purely to capture the inventory dict for /health;
  load() is idempotent (refresh=True) and cheap, so this is deliberate,
  not a leftover.

Test plan

• ruff check agentdeck/ tests/
• ty check agentdeck
• pytest tests/ -q (added test_open_close_lifecycle: open → session_for
  plumbing (no live model, SQLite fallback) → aclose, then a second
  aclose() to confirm idempotency)

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.