Summary

Some custom providers return SSE streams even when stream=False is not honored. This caused title_generation and other auxiliary tasks to fail with errors like:

Auxiliary title_generation: LLM returned invalid response (type=str): 
'data: {"id": "chatcmpl-1b5552757d54", "object": "chat.completion.chunk"...'
Expected object with .choices[0].message

Changes

• Add _aggregate_stream_response() — consumes SSE chunks into a standard .choices[0].message completion object
• Detect streaming responses in _validate_llm_response():
  • Raw SSE strings containing data: prefix
  • Iterable objects without .choices attribute
• Support OpenAI SDK ChatCompletionChunk objects and raw SSE text lines
• Preserve reasoning_content from thinking models (e.g. Qwen, DeepSeek)
• Add comprehensive test suite (13 tests)

How it works

When a custom provider ignores stream=False and returns SSE data:

1. _validate_llm_response detects the response is a stream (string with data: or iterable without .choices)
2. Transparently calls _aggregate_stream_response to consume all chunks
3. Returns a SimpleNamespace matching the expected .choices[0].message shape
4. Downstream code works unchanged

Test Plan

• Unit tests pass (13/13)
• Verified with custom provider that returns SSE by default
• title_generation now succeeds instead of failing

Fixes #7264