[FEATURE] Improve computer vision for all models

[edenreich]

Summary

Look into https://github.com/AmrDab/clawdcursor

Which will allow to classify screenshots and check if that works well with deepseek.

It should be way faster instead of relying on Opus models or any vision models really.

Decoupling vision from the models is the right architectural decision for computer use.

Acceptance Criteria

• Computer use is does not rely on the vendor specific vision
• There is a clear path toward implementing it
• A follow up ticket is created with the clear path (if applicable) - in case it's not possible to work with the above mentioned open source project etc, all of these are documented in a follow up