Race condition panic in streaming output detection

[ckavili]

Describe the bug

Streaming chat completions with output detection causes a panic due to a race condition between concurrent tasks. The completion_state.completions.get(&choice_index).unwrap() call fails because the spawned task hasn't inserted the entry before the main task tries to read it.

Platform

• Orchestrator version: 0.16.0 (commit a220c15)
• Rust version: (from container image)
• Environment: OpenShift, detectors running in same cluster (~8ms response time)

Sample Code

Request payload that triggers the bug:

{
  "model": "llama32",
  "messages": [{"role": "user", "content": "hello"}],
  "stream": true,
  "detectors": {
    "input": {"hap": {}, "language_detection": {}, "prompt_injection": {}},
    "output": {"hap": {}, "language_detection": {}}
  }
}

Expected behavior

Streaming response with output detection completes without error.

Observed behavior

Panic on first request:

thread 'tokio-runtime-worker' panicked at src/orchestrator/handlers/chat_completions_detection/streaming.rs:573:76:
called `Option::unwrap()` on a `None` value

Happens consistently, not intermittently. Disabling output detectors fixes it.

Additional context

Race is between:

• process_chat_completion_stream (spawned): inserts into completion_state.completions
• process_detection_batch_stream (main): reads from completion_state.completions

Affected lines: streaming.rs:506 and streaming.rs:589

Workaround: use stream: false or disable output detectors.