[Feature request] Automatic vision fallback — describe images with an auxiliary vision model

[Johell1NS]

Proposal: when the bot's active model doesn't support image input, automatically describe the image with a user-configurable auxiliary vision model and forward the description (plus the original caption) to the main model as text.

Why: today, if the model has no vision support, the photo is discarded and the user only gets "model doesn't support image input" (plus the caption at best). The visual content is completely lost. This feature:

• eliminates that information loss, regardless of which model is running;
• lets users pair a cheap/fast vision model (for describing) with a more capable non-vision model (for reasoning), instead of being forced to use an expensive vision model for everything.

Design principles: opt-in and backward compatible — if the fallback model isn't configured, current behavior is unchanged.

Detailed request: the full proposal (requirements, prior art from Hermes Agent and OpenClaw) is tracked in issue #151.

Status: the behavior described above is still present in main. The codebase has since matured exactly where this feature needs it — model-capabilities-service (getModelCapabilities/supportsInput), model-selection-service (getStoredModel), and the runtime /settings command (#171) — so it could now be implemented in a fairly surgical way.

Question to @grinev: is this a direction you'd be willing to accept? If so, I'm happy to contribute in small reviewable steps (e.g. starting with photo-handler only, no new commands), or to provide whatever input you'd find useful. Thanks!