Security/stability: cap LLM response bodies before reading into memory

[CryptoJones]

Problem

OpenAiCompatibleLlmClient reads successful LLM responses with ReadAsStringAsync, and the error path reads the entire body before truncating it for logs. A malicious, misconfigured, or broken endpoint can return an arbitrarily large body.

Impact

A single draft request can force large memory allocations and destabilize the API process.

References

• api/ApplyTrack.Api/Llm/OpenAiCompatibleLlmClient.cs

Acceptance criteria

• Use HttpCompletionOption.ResponseHeadersRead.
• Reject responses with Content-Length over a configured/default cap.
• Stream response bodies with a hard byte cap before JSON parsing/logging.
• Tests cover oversized success and error responses.