fix(providers): add circuit breaker for Responses API fallback

[mohamed-elkholy95]

Summary

• Add a proper circuit breaker for Responses API fallback in OpenAICompatProvider
• After 3 consecutive compatibility errors for a (model, reasoning_effort) key, skip the Responses API and go straight to Chat Completions
• Circuit probes again after 5 minutes (half-open state) — recovery is automatic
• Success immediately resets the failure counter
• Addresses the issue where a transient Responses API outage could permanently degrade output for the lifetime of the process

Test plan

• 7 new tests covering: default availability, threshold opening, model isolation, success reset, time-based probe, below-threshold pass, reasoning_effort key separation
• All 209 existing provider tests pass (no regressions)
• ruff check --select F401,F841 clean

🤖 Generated with Claude Code

[chengyongru]

Review: fix(providers): add circuit breaker for Responses API fallback

结论: Approve

这是一个设计良好的断路器实现。

优点：

• 按 (model, reasoning_effort) 键隔离，一个模型的故障不影响其他模型
• 阈值 (3次) 和探测间隔 (5分钟) 作为模块级常量清晰定义
• 半开 (half-open) 状态设计正确：超过冷却期后允许一次探测
• 成功时立即重置所有计数器，恢复迅速
• 在非流式和流式两条代码路径都正确集成了 _record_responses_success/failure
• 测试覆盖充分 (7个测试)，覆盖了所有关键场景

小建议（不阻塞合并）：

1. _record_responses_failure 里有内联 from loguru import logger，建议移到文件顶部与其他 import 保持一致。这不是热路径，每次调用都走 import 缓存不会有性能问题，但放在顶部更符合惯例。

2. 断路器的 _responses_failures 和 _responses_tripped_at 字典会随着不同 key 不断增长（虽然实际上 key 空间很小）。如果在意极端情况，可以在 _record_responses_success 中确认两个 dict 的清理。当前实现已经在做 pop 了，所以这已经处理了。

Good to merge.

[chengyongru]

Review: fix(providers): add circuit breaker for Responses API fallback

Verdict: Approve

Well-designed circuit breaker implementation.

Pros:

• Isolated per (model, reasoning_effort) key — one model's failures don't affect others
• Threshold (3) and probe interval (5 min) are clearly defined as module-level constants
• Correct half-open state: allows one probe attempt after the cooldown period
• Success immediately resets all counters for quick recovery
• Properly integrated in both non-streaming and streaming code paths
• Adequate test coverage (7 tests) covering all key scenarios

Minor suggestions (non-blocking):

1. _record_responses_failure has an inline from loguru import logger. Consider moving it to the top of the file with the other imports for consistency. This isn't a hot path so the import cache handles it fine, but top-level is more conventional.

2. The _responses_failures and _responses_tripped_at dicts grow with distinct keys (though the key space is tiny in practice). _record_responses_success already does pop to clean up, so this is handled.

Good to merge.