Archive failed inference sessions through discard lifecycle

llm-bridge: consume published pi-ai, file-backed OAuth, dynamic Anthropic discovery + prompt caching

i think code is clean enough atm