Feature: Circuit breaker for model rate limiting in Ralph

[tamirdresher]

Summary

Adds a circuit breaker pattern for automatic model fallback when Copilot rate limits are hit during Ralph operations.

Closes #450

Problem

When running multiple Ralph instances, premium model rate limits cause cascading failures — all Ralphs go down at once.

Solution

Classic circuit breaker (Hystrix/Polly pattern) applied to model selection:

• CLOSED (normal): Use preferred model (default: \claude-sonnet-4.6)
• OPEN (rate limited): Fall back through free-tier chain: \gpt-5.4-mini\ → \gpt-5-mini\ → \gpt-4.1\
• HALF-OPEN (testing): After cooldown, try preferred model. 2 consecutive successes → close circuit.

What's included

• .squad-templates/ralph-circuit-breaker.md\ — Full documentation with:
  • Pattern explanation and state diagram
  • PowerShell functions (\Get-CurrentModel, \Update-CircuitBreakerOnSuccess, \Update-CircuitBreakerOnRateLimit)
  • State file schema (.squad/ralph-circuit-breaker.json)
  • Integration instructions for
    alph-watch.ps1\
  • Metrics tracking (fallbacks, recoveries)

Key design decisions

• Free-tier models (0x multiplier) as fallbacks — unlimited usage, no quota burn
• Cooldown-based recovery (default 10 min) — avoids hammering a rate-limited endpoint
• Persistent state file — survives Ralph restarts
• Cascading fallback chain — if even the first fallback is limited, tries the next

Testing

Validated with 8 concurrent Ralphs across 8 repos. Rate limit hit after ~2 hours of intensive polling. Circuit breaker engaged, Ralphs continued on \gpt-5.4-mini, recovered to preferred model after 10-minute cooldown.