[Dhirenderchoudhary]

Builds on #430 (and #429). Adds image and video to the Workflows canvas so a flow can mix text, audio, image and video models.

New nodes:

• Image input and Video input. A video is turned into N evenly spaced frames in the browser (default 4), shown as thumbnails.
• Image generation (text -> image) through /v1/images/generations, model_type = image_generation.
• Vision (image or frames -> text) through /v1/chat/completions with OpenAI image content, for vision models such as moondream or llava.
• Image output.

New templates: "Image from a description", "Describe a video", "Spoken image request" (mic -> STT -> image). The plain-English planner picks the new nodes up from the registry, the keyword fallback knows image/video requests, and the readiness catalog recommends sd-turbo and moondream. Vision models now rank after text models in the LLM dropdowns so a chat node doesn't default to moondream.

Testing:

• 66 workflow tests (new: port types, image edges, request bodies, frame timing, planner), full dashboard suite 485 passed, build, eslint and tsc clean on changed files.
• Ran all three templates in Chromium against a local stack (no GPU): sd-turbo behind a small OpenAI-compatible image server and moondream on Ollama, both registered as deployments. All calls 200.
• On laptop CPU, image generation took ~46 s warm while another model was using the CPU, and moondream's description of 4 frames was weak. These are CPU numbers; on a GPU sd-turbo is a single step.

[Priyanshu-u07]

Merged main into this branch (b8a6418) to resolve the conflicts after #430. Your design is unchanged; only main's two fixes are carried over: the recorder fix (now on setInput) and the run-in-progress toast. Full dashboard suite: 488 passed, build passes. Please pull before pushing again.

Two follow-ups for you, not blocking:

• Image generation Deploy uses engine: "generic", which neither worker pools nor Kubernetes can run. Added to Workflow Deploy offers engines the pools can't run (speaches, generic) #433.
• Planner keyword rules need word boundaries: "voice assistant that helps customers withdraw cash" builds mic → image generation ("withdraw" contains "draw"), and "speech recognition" goes to the vision pipeline.