fix(workflow): controller readiness, atomic orchestrator swap, and shutdown cleanup

[shoom1]

Addresses review finding 4 (workflow readiness and lifecycle ownership). Verified all claims against the code first; every one held, plus one unnamed race.

Problems (confirmed)

• is_ready meant "manager object exists": _background_init published self._workflow before initialize_services() ran, so a failed or still-initializing manager was reported ready, and ensure_initialized() returned True before ever consulting _init_error.
• Unnamed extra: initialize_services() guarded with check-then-act (if self._initialized: return) and no lock — background init racing a first user message (process() → _ensure_initialized()) ran the whole body twice (duplicate registry refresh, duplicate service creation).
• Orchestrator swap assigned the new manager before initializing it: a failed swap left a broken manager with the working one lost; a successful swap orphaned the old manager without cleanup() (which tears down real resources — sessions, sandbox manager).
• App shutdown (background_init CM exit) only cancelled the init task; manager.cleanup() was never called anywhere in the app.

Fix

• Publish-after-success in _background_init; failure/cancel paths clean up the partially-built manager best-effort. is_ready/ensure_initialized become truthful with no new state to drift (states remain derivable: uninitialized / initializing / ready / failed).
• asyncio.Lock around initialize_services() in BaseWorkflowManager — concurrency-safe and idempotent for both orchestrators and any direct library consumer.
• Swap: initialize replacement → swap atomically → clean up old; on init failure the working manager stays in place and the broken replacement is cleaned up.
• WorkflowController.close(): idempotent (cancel init → detach → cleanup), wired into the background_init CM finally, which is the app's shutdown path (app.py:555).

Deliberately not an explicit state enum: the reviewer suggested one, but with publish-after-success the four states are already derivable from (_workflow, _init_task, _init_error) — an enum would duplicate state that can drift.

Tests

TDD: 9 new tests written first and watched fail — readiness during in-flight init, failed-init cleanup + not-ready, ensure_initialized blocking until services finish, swap-failure-keeps-old / swap-success-cleans-old, idempotent close(), CM-exit cleanup, and concurrent + sequential initialize_services single-run. Full offline suite: 1990 passed. (The traitlets RuntimeWarning visible in full runs was verified pre-existing via stashed A/B run — order/GC-dependent, in untouched sandbox tests.)

https://claude.ai/code/session_01SwkKXQpy3JLEqrPkQA8QRv