Add optional TwelveLabs video segmentation backend (Marengo + Pegasus)

[mohit-twelvelabs]

Hi! I'm Mohit, I work at TwelveLabs (@mohit-twelvelabs).

What this adds

An opt-in segmentation backend that analyzes the video frames and audio directly via TwelveLabs, as an alternative to the default Whisper-transcript + LLM pipeline:

• Marengo (search) — semantic retrieval to locate clips relevant to the user's prompt.
• Pegasus (analyze) — generates the per-segment summary and tags.

It returns the exact same {start, end, summary, tags} structure the rest of the app already consumes, so the analysis table, re-analyze tab, clipping, and export flows all work unchanged.

Why it helps PreenCut

The current pipeline keys off speech (Whisper → transcript → LLM). For speech-light footage — sports, b-roll, surveillance, music videos, silent demos — there's little transcript to segment on. Segmenting straight from the visual + audio content covers that gap.

Opt-in / non-breaking

Disabled by default. The default SEGMENTER_TYPE=llm path is untouched. To enable:

export SEGMENTER_TYPE=twelvelabs
export TWELVELABS_API_KEY=your_key
# optional: TWELVELABS_INDEX_NAME, TWELVELABS_MAX_CLIPS

New config lives alongside the existing LLM_MODEL_OPTIONS in config.py; the processor (modules/twelvelabs_processor.py) mirrors LLMProcessor.segment_video's contract and is swapped in processing_queue.py only when opted in. twelvelabs>=1.2.8 added to requirements.txt.

How it was tested

• No-network unit tests for the JSON/segment parsing and the overlap filter (tests/test_twelvelabs_processor.py), runnable as plain Python or via pytest.
• A live smoke test (gated on TWELVELABS_API_KEY, skipped when unset) confirming a Marengo text embedding returns a vector.
• Verified live against the SDK that Pegasus analyze returns well-formed {start, end, summary, tags} segments and Marengo search returns relevant time windows that the overlap filter applies correctly. The slow upload/index step uses the documented SDK signatures (tasks.create / wait_for_done).

You can grab a free API key at https://twelvelabs.io — there's a generous free tier.