Parent Epic

stacklok/stacklok-epics#201

Depends on: #3733

Context

The epic acceptance criteria requires that the optimizer integrates with the circuit breaker functionality so that unhealthy backends don't cause complete failures. vMCP already has a health monitoring system (pkg/vmcp/health/) that tracks backend health. When a backend is unhealthy at session creation time, its tools should be excluded from that session's optimizer.

This issue does not introduce within-session health transitions or capability refresh logic. Health status is evaluated once at session creation — if a backend is unhealthy when the session starts, its tools are excluded for the lifetime of that session.

Requirements

1. When a backend is unhealthy at session creation time, its tools must be excluded from that session's FindTool results.
2. When a backend is healthy at session creation time, its tools are included normally.
3. Create acceptance tests that validate this behavior.

High-Level Implementation

• Investigate how the health monitor (pkg/vmcp/health/) currently communicates backend health status
• Ensure the optimizer's per-session tool scoping (from Back the dummyOptimizer with a shared SQLite instance and basic similarity search #3731) respects backend health at session creation time — unhealthy backends' tools are simply not passed to the optimizer factory
• Write acceptance tests that:
  • Start with multiple backends, some healthy and some unhealthy
  • Create a session and verify FindTool only returns tools from healthy backends
  • Create another session after a previously unhealthy backend has recovered and verify its tools are now included

Acceptance Tests

• Unhealthy backend excluded: When a backend is unhealthy at session creation, its tools are not returned by FindTool for that session
• Healthy backend tools unaffected: Tools from healthy backends are returned normally
• New session sees recovered backend: A session created after a backend recovers includes that backend's tools
• No crashes: The optimizer does not crash or hang when backends are unhealthy at session creation time