Add agent health monitoring and graceful shutdown support

[97nitt]

Summary

Phase 2 agent-side changes for the egress agent pull model improvements — goodbye event handling, graceful shutdown, heartbeat during backpressure, and instance ID tracking.

Goodbye Event Handling:

• EventsClient handles new goodbye SSE event type from orchestrator
• BaseEgressAgentService._handle_goodbye triggers graceful shutdown on goodbye

Graceful Shutdown:

• _trigger_graceful_shutdown: notifies orchestrator, stops threads, signals process to exit
• register_signal_handlers: registers SIGTERM/SIGINT handlers (called from gunicorn post_worker_init hook or local dev main)
• _shutting_down flag prevents double shutdown when both goodbye and signal paths fire
• Under gunicorn, signals master process so all workers shut down
• SSE receiver and ops_runner/results_publisher threads run as daemon threads so they don't block process exit

Backpressure Heartbeat:

• OperationsPoller sends heartbeat to orchestrator during backpressure to maintain activity tracking
• BackendClient.send_heartbeat: POST /api/v1/agent/heartbeat

Shutdown Notification:

• BackendClient.notify_shutdown: POST /api/v1/agent/shutdown (15s timeout)

Instance ID (Phase 3 prep):

• BackendClient generates a UUID instance ID on construction
• All requests include x-mcd-agent-instance-id header (orchestrator ignores it for now)
• SSE connection includes the header via extra_headers on SSEClientReceiver
• Instance ID logged at startup

Other:

• Fixed misleading "Woken by work_available notification" log during shutdown (was actually woken by stop())

Test plan

• Unit tests pass
• Integration tested on dev (hermes-agent, sna-artemis-agent)
• Verify goodbye flow: orchestrator sends goodbye → agent logs shutdown → process exits
• Verify SIGTERM flow: kill -TERM <pid> → agent notifies orchestrator → process exits
• Verify heartbeat during backpressure: all worker threads busy → agent still sends heartbeats

Dependencies

• Orchestrator changes: monte-carlo-data/data-collector#2244
• SSE reconnect fix: Fix SSE reconnect race that disables heartbeat checker #14 (merged)

[mrostan]

lgtm, left a nit

[97nitt]

Code Review Summary

0 blockers, 2 issues, 3 suggestions, 1 nit — reviewed by correctness, security, devdocs, testing, architecture agents.

Fixed (2bed82d)

• F1 [ISSUE] _trigger_graceful_shutdown double-execution guard was not thread-safe — replaced _shutting_down boolean with threading.Lock
• F2 [ISSUE] Missing tests for BackendClient.send_heartbeat and notify_shutdown — added test_backend_client.py with 7 tests
• F3 [SUGGESTION] notify_shutdown docstring said "Best-effort" but raises on failure — fixed docstring
• F5 [SUGGESTION] Missing test for heartbeat during backpressure — added test_backpressure_sends_heartbeat

Skipped

• F4 [SUGGESTION] Backpressure heartbeat frequency — activity timeout (120s) is 2x poll interval (60s), sufficient margin
• F6 [NIT] _headers(**extra: str) type annotation — correct Python, left as-is