- Propagated llama.cpp /props input modalities through selected-model runtime refresh.
- Added a regression test for cached text-only local vision models becoming image-capable after refresh.
- Updated the coding-agent changelog for the local vision detection fix.
Fixes#4654