Scene Edit improvements

[philpax]

• Prompt-cache everything up to the user's actual input
• Consider using Gemma 4 for inputs (wait for model inference + templating to stabilise)
  • Also see if transformers is good enough for this It was not. There's no built-in way to shift the embeddings to the CPU, which defeats the purpose of using a per-layer embeddings model. We will be on llama-cpp for the indefinite future.
• Persistent scene edits: keep an ongoing session open, so that new prompts have the context of previous prompts available. Try with and without images remaining in context.
• Raw filtering: Option to disable the prompt enhancement, so that what you type is what you get. A VLM would still be necessary to do filtering, and any safety rewriting would be promoted to total cancellation.
• "Repaint scene": hypnagogia-style repainting, where the scene is analysed by the VLM, and then the frame is repainted in place with the VLM's prompt. This can stabilise the experience and/or convert vaguely-defined features into actual objects (e.g. door-shaped objects into doors). Separate keybind?
• "Add something interesting": a button that looks at the current scene and adds something interesting. Could be periodic, could be whenever the user stops issuing any inputs for a while.