feat: 新增可选的 TwelveLabs 供应商（Pegasus 视频理解 + Marengo 多模态向量）/ Add opt-in TwelveLabs provider - #15

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

简介（中文）

为项目新增一个 完全可选 的 TwelveLabs 视频理解供应商，沿用现有的 provider 工厂模式，不改变任何默认行为——不配置 TwelveLabs API Key 时系统一切照旧。

• Pegasus（视频理解 / analyze）：对成片或素材做内容分析，可用于自动生成分镜描述、镜头摘要、内容审核等创作辅助。
• Marengo（多模态向量 / embed）：为视频与文本生成 512 维向量，可用于素材库语义检索、画布节点之间的素材匹配与去重。

配置方式与其它供应商一致：在“API 密钥管理”页面新增一条 provider = twelvelabs 的密钥即可。免费额度可在 https://twelvelabs.io 获取。

What this adds (English)

A new opt-in TwelveLabs provider that plugs into the existing ProviderFactory pattern, mirroring VectorEngineProvider (httpx-based, same log_provider_call decorator and Chinese docstrings/logging style):

• backend/src/services/provider/twelvelabs_provider.py
  • analyze(...) — Pegasus video understanding (accepts a public video_url or an uploaded asset_id; Pegasus 1.5 does not accept a bare video_id).
  • embed_text(...) / embed_video(...) — Marengo 512-dim multimodal embeddings.
• Wired into ProviderFactory.create("twelvelabs", ...), the provider validation whitelist + APIKeyProvider enum, and APIKeyService.get_models (pegasus1.5 for video, marengo3.0 for embeddings).
• Frontend: twelvelabs option added to the API Key management dropdown / label / icon maps.

Why it helps this project

AICON is a node-based AI video creation workflow. Pegasus gives the canvas a way to understand generated/imported clips (auto shot descriptions, summaries, moderation), and Marengo gives semantic asset retrieval/dedup across canvas nodes — both natural extensions of the existing storyboard/asset pipeline.

Opt-in / non-breaking

This is purely additive. No defaults change; if no TwelveLabs key is configured, behavior is identical to before. It reads its own API key from the repo's existing per-user, DB-backed API Key management (encrypted) — no key is hardcoded.

How it was tested

• No-network unit tests for the factory wiring, the multipart embed encoding, the Pegasus payload shape, and the embedding extraction (backend/tests/unit/test_twelvelabs_provider.py) — 6 passed.
• A live integration test gated on TWELVELABS_API_KEY (skipped when unset). I ran it against the real API and confirmed a Marengo text embedding returns a 512-dim vector — passed.
• I live-verified both REST contracts against api.twelvelabs.io/v1.3: Marengo marengo3.0 requires multipart/form-data (even for text-only; raw vector key is float), and Pegasus pegasus1.5 accepts video: {type, url} with max_tokens >= 512.
• Ran black / isort on the new files only (kept the diff minimal — existing files were edited surgically, not reformatted).

Note: my local macOS sandbox couldn't load libmagic (a pre-existing dependency of an unrelated conftest import chain), so I ran the focused TwelveLabs tests with a tiny magic stub injected; they pass cleanly. The full suite should run normally in your CI/Linux env where libmagic is present.

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