feat(resilience): add circuit breaker pattern for production stability

[lsvishaal]

PR Description for the Circuit Breaker

---

Summary

Implements a production-grade Circuit Breaker pattern to prevent cascade failures when external services (LLMs, APIs) are degraded or unavailable.

Why This Change?

Bindu agents call external LLMs (OpenAI, Anthropic, etc.) which can experience:

• Rate limits
• Latency spikes
• Complete outages

The existing retry mechanism (bindu/utils/retry.py) handles transient failures but keeps retrying indefinitely on degraded services. This can:

• Block workers waiting on slow/dead services
• Exhaust connection pools
• Cause cascade failures across the agent network

Circuit breaker complements retry — retry = "try again" wheras circuit breaker = "stop trying, it's broken."

What's Included

Component	Description
bindu/utils/circuit_breaker.py	Core implementation (580 lines)
bindu/settings.py	CircuitBreakerSettings config class
tests/unit/test_circuit_breaker.py	34 async tests (TDD approach)
examples/circuit_breaker_demo.py	Manual verification script

Key Features

• Three-state FSM: CLOSED → OPEN → HALF_OPEN → CLOSED
• Async-native: Thread-safe state transitions with asyncio.Lock
• Three usage patterns: Decorator, context manager, direct call
• Global registry: Monitor all circuit breakers in one place
• Configurable: Via CircuitBreakerSettings (Pydantic)
• Observable: Sentry integration for state change alerts

Usage Example

from bindu.utils.circuit_breaker import circuit_breaker

@circuit_breaker(name="openai", failure_threshold=5)
async def call_openai(prompt: str) -> str:
    # After 5 consecutive failures, circuit opens
    # Calls fail-fast until recovery timeout
    ...

Testing

# Run circuit breaker tests
pytest tests/unit/test_circuit_breaker.py -v

# Run demo
python examples/circuit_breaker_demo.py

# Verify no regressions (516 tests)
pytest tests/ -q

Reference

Based on Michael T. Nygard's "Release It!" pattern for production resilience.

✅ Checklist

• Tested thoroughly (34 tests, all passing)
• Followed project style and linting (ruff clean)
• Added demo script for verification
• No regressions (516 existing tests still pass)