Guard repeated graph-memory LLM failures

[junyangyan4-svg]

Summary

• add LlmFailureGuard for permanent graph-memory LLM API failures
• pause repeated calls after permanent 4xx responses such as 403/404/422
• reset the guard immediately after a successful LLM call

Why

This turns a runaway failure mode into a bounded cooldown when the configured memory LLM is misconfigured or unavailable.

Split from #39.

Verification

• npm test
• npx tsc --noEmit

[adoresever]

核心方案已采纳并合并：连续配置类错误的冷却保护是必要的。合并时把触发范围收窄为持久配置错误（401/403/404），避免 400/422 等请求级错误错误地停用 LLM，并补充了测试。感谢贡献。