v0.12.0 — data_analysis skill + media download + local image backend

Rounds out Chimera's media + data capabilities, all in the honest orchestrate, don't reimplement spirit.

Added

• data_analysis skill — names how an agent actually "does ML": writes a self-contained pandas + scikit-learn script (load → explore → model → evaluate) and runs it in the execute_code sandbox. Orchestration, not a reimplementation of sklearn.
• download_media tool — download a video (or just its audio, mp3) from YouTube + 1000+ sites into the workspace. Wraps yt-dlp (not pytube — single-site and perpetually broken), so it survives player/cipher/age-gate churn. Opt-in media-dl extra; audio extraction needs ffmpeg. Pairs with transcribe_audio.
• Local image backend for generate_image — set CHIMERA_IMAGE_BACKEND=local to run FLUX.1-schnell (Apache-2.0) via diffusers fully offline (opt-in, GPU-heavy imagegen-local extra). Hosted OpenAI stays the default.

Honest scope

Studied pytube / CogVideo / OpenCV: pytube → wrapped yt-dlp (more robust); CogVideo (video generation) deliberately not vendored — a heavyweight trained model, hosted-API territory if ever needed; OpenCV needs no dedicated tool — the agent already import cv2s in the sandbox. Chimera runs diffusion/Whisper models; it does not train them.

Gate: ruff clean, 1215 passed / 2 skipped, mypy clean.

PyPI: pip install chimera-agent==0.12.0