Fix clip_timestamps format, Windows UTF-8 console, and non-dict LLM response

[Ahmed-Ezzat20]

Summary

Three bug fixes discovered during end-to-end testing of the pipeline:

1. clip_timestamps format bug (faster-whisper backend)

_transcribe_faster_whisper() passes speech_clips_sec as a list of dicts ([{"start": 0.5, "end": 3.2}, ...]) to clip_timestamps, but BatchedInferencePipeline.transcribe() expects a flat list of seconds ([0.5, 3.2, ...]). This causes a TypeError at runtime.

Fix: Convert to flat [start1, end1, start2, end2, ...] format.

2. Windows console UnicodeEncodeError

On Windows, sys.stdout defaults to cp1252 encoding, which cannot represent Arabic, CJK, or other non-Latin characters. Any print() or log message containing a non-Latin project slug (e.g. مين-هو-مستر-عزت) crashes with UnicodeEncodeError.

Fix: Reconfigure stdout/stderr to UTF-8 with errors="replace" at CLI entry point, guarded by sys.platform == "win32".

3. describe.py crash on non-dict LLM response

json_repair.loads() can return a list, string, or None when small/local LLMs (e.g. Ollama qwen3:4b) produce malformed JSON. The subsequent .get() call crashes with AttributeError: 'list' object has no attribute 'get'.

Fix: Guard with isinstance(description, dict) check, wrapping non-dict responses into the expected schema.

Test plan

• Verified all modified modules import correctly
• Test faster-whisper transcription with --method faster-whisper (requires GPU or CPU with the model)
• Test pipeline with non-Latin project names on Windows
• Test describe stage with a small local LLM (e.g. Ollama qwen3:4b)