add optional twelvelabs pegasus transcription backend

[mohit-twelvelabs]

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

This PR adds an optional TwelveLabs Pegasus transcription backend alongside the default Whisper pipeline.

What it adds

• A new decipher/pegasus.py module with video_to_srt(...): uploads the video as a TwelveLabs asset, waits for it to be ready, and asks Pegasus to return SubRip (SRT) subtitles (output is fence-stripped and validated against an SRT cue regex).
• A --backend {whisper,pegasus} flag on decipher transcribe (default whisper), wired through action.transcribe(...) via a new backend argument.
• An optional dependency group: pip install "decipher[pegasus]" (twelvelabs>=1.2.8).
• README docs and a tests/test_pegasus.py (no-network unit tests for parsing/validation + a key-gated wiring test that skips without TWELVELABS_API_KEY).

Why it helps

Pegasus is a video-understanding model that reads the video directly rather than just the extracted audio track, so it can use on-screen context (overlaid text, speaker cues) when transcribing. It also runs in the cloud, so users without a local GPU get an alternative to Whisper. Marengo/Pegasus and Whisper are complementary, not a replacement.

Opt-in / non-breaking

Default behaviour is unchanged: backend defaults to whisper, the twelvelabs package is only imported inside the pegasus branch, and it's an extras_require (not installed by default). Existing Whisper users see no difference.

How it was tested

• tests/test_pegasus.py passes (5 passed): fence stripping, SRT cue regex, missing-key error, and SDK construction with a key.
• End-to-end verified against the live TwelveLabs API: a short clip was uploaded, polled to ready, analysed with Pegasus 1.5, and returned a valid SRT cue (~8s). Marengo text embeddings also confirmed (512-dim) while validating the SDK contract.
• Formatted with black.

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