feat(ai-service): export Prometheus metrics for circuit breaker (closes #134)

[Moonwalker-rgb]

What

Closes #134 — adds Prometheus metrics for the AI service circuit breaker so unhealthy provider trips are observable in dashboards rather than from user-visible degradation.

Changes

• Adds three Prometheus instruments updated on every breaker transition:

Metric	Type	Description
circuit_breaker_state	Gauge	Current state (0=CLOSED, 1=HALF_OPEN, 2=OPEN), labeled by breaker_name
circuit_breaker_failure_count_total	Counter	Cumulative failures per breaker
circuit_breaker_recovery_time_seconds	Histogram	Time spent OPEN before transitioning to HALF_OPEN

• Metric updates happen inside the existing _lock that already guards breaker state, so the exported values can never diverge from the underlying state.
• The initial state (CLOSED) is published in __init__ so every instantiated breaker appears in the gauge even before any traffic flows.
• Exports CIRCUIT_STATE_CLOSED|HALF_OPEN|OPEN constants and a set_circuit_state(...) helper to keep call sites readable.

Tests

• New TestCircuitBreakerMetrics class with five focused tests:
  • Initial state published
  • Failure increments counter / gauge flips on threshold
  • OPEN → HALF_OPEN updates histogram and gauge
  • HALF_OPEN → CLOSED via success
  • HALF_OPEN → OPEN via failure
• All 9 tests in tests/test_circuit_breaker.py pass locally in ~1.5s.

Compatibility

• Additive only. Existing CircuitBreaker callers are untouched.
• Uses the existing metrics.py registry; no new dependencies, no new registry.

Notes

• circuit_breaker_failure_count_total is cumulative since the last reset, matching Prometheus counter semantics. Operators should compare rate over a window rather than the absolute value.
• The histogram uses prometheus_client default buckets. If finer-grained slices are needed later, pass buckets=[1, 5, 10, 30, 60, 120].

[Moonwalker-rgb]

cc @gbengaeben @CodeMayor — recent contributors to app/ai-service and project tooling. Could one of you (or another ai-service owner) take a first pass? Specifically curious whether the gauge encoding (0/1/2) and the cumulative counter semantics fit how app/backend already scrapes these metrics.

[kilodesodiq-arch]

LGTM