AbstractVision v0.3.7

• AbstractCore plugin: add explicit local model residency control for in-process diffusers, mflux, and sdcpp backends through load_resident_model(...), list_loaded_models(...), list_resident_models(...), and unload_resident_model(...), with load_model(...) / unload_model(...) compatibility aliases for Core route adapters.
• Residency semantics: preserve explicit resident models across model switches while keeping the previous unload-on-switch behavior for non-resident request-warm backends, and report stable process-local load_id / backend_kind / resident metadata instead of relying on backend private fields.
• Safety/robustness: reject OpenAI/OpenAI-compatible HTTP backends for residency control even on localhost, normalize provider/task filters consistently, reject ambiguous unload requests, and keep unload behavior deterministic for injected local backends.
• Tests/docs: expand AbstractCore plugin coverage for local preload/list/unload flows, request-warm loaded-state reporting, task-aware loaded-model filters, injected backend residency, and switch-survival behavior; document the new local residency control surface and regenerate llms-full.txt.