feat: add opt-in TwelveLabs Pegasus on-screen visual context

[mohit-twelvelabs]

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

---

简体中文（摘要）

目前的总结 / 术语提取步骤（_4_1_summarize.py）只依据音频转写文本，因此看不到画面上的文字、产品 / 品牌名称、UI 标签和图表。本 PR 新增一个可选的 TwelveLabs Pegasus 视觉上下文：开启后，Pegasus 会对视频画面分析一次，把"画面里有什么"的描述注入到总结 / 术语提示词中，从而帮助消除画面内容在断句与翻译时的歧义（例如专有名词、屏幕上出现的品牌名）。

完全可选、不破坏现有行为：默认关闭（pegasus.enabled: false），未配置 key 时流程与之前完全一致。可在 twelvelabs.io 免费获取 API key，有慷慨的免费额度。

---

What this adds

VideoLingo's summary/terminology step (core/_4_1_summarize.py) builds the theme and glossary from the transcript only, so it's blind to on-screen text, product/brand names, UI labels, and charts that the narration never says aloud. This PR adds an opt-in TwelveLabs Pegasus visual-context pass:

• New core/utils/pegasus_context.py — uploads the input video once and asks Pegasus 1.5 to describe the on-screen visual layer.
• The description is injected into get_summary_prompt(...) (new optional visual_context arg) so the summary + terminology extraction can disambiguate proper nouns and on-screen terms before translation/segmentation.
• New pegasus: block in config.yaml.

Why it helps VideoLingo

The pipeline's translation quality hinges on the summary/terminology step. Feeding it what's actually on screen (e.g. a brand logo, an app name in the UI, a chart label) lets it pick the right translation for ambiguous terms and proper nouns that the audio alone can't pin down.

Opt-in / non-breaking

• Disabled by default (pegasus.enabled: false). With no key configured, get_visual_context() returns "" and the pipeline behaves exactly as before.
• Key is read from pegasus.api_key, falling back to the TWELVELABS_API_KEY env var — no key is hardcoded.
• All Pegasus errors are caught with a warning so a hiccup never breaks translation. Result is cached to output/log/ for resumed runs. >200MB direct-upload cap is guarded.

How it was tested

• tests/test_pegasus_context.py: 4 no-network unit tests verifying the feature is genuinely opt-in (disabled / no-key → no-op) and that the prompt is unchanged without context but injects it when present — all pass.
• A live test gated on TWELVELABS_API_KEY (skipped without it) that performs a real asset upload + Pegasus analyze — verified passing locally against the TwelveLabs API (returns a non-empty on-screen description).
• Modules compile and import cleanly; config loads.

Note: I couldn't run VideoLingo's full GPU/WhisperX pipeline end-to-end in my Linux sandbox (heavy CUDA/ML deps), so the integration follows the repo's existing backend conventions (requests-style modules, load_key, rprint, output/log/ caching) and was validated at the unit + live-API level. Happy to adjust to your preferences.

twelvelabs>=1.2.8 is added to requirements.txt (pure-Python SDK; the import is guarded so it's only needed when the feature is enabled).

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