bug: model override registry leaks across model directories

Problem

ModelRegistry keeps override state in a mutable global map and does not clear provider overrides when a later model directory has no override file. This makes custom model registries sticky across unrelated NetLogo models in the same JVM session.

Current behavior

• ModelRegistry.loadOverride(modelDir) merges override models into the shared providers map.
• If the next model directory has no models-override.yaml, the method returns None and leaves the prior merged state intact.
• There is no reset-to-bundled step when switching models or loading config from a directory without overrides.

Why this matters

This can leak one model’s custom registry into another model that never opted into it:

• llm:list-models can show stale custom models
• llm:set-model validation can succeed for models that should not exist in the current model directory
• cross-model testing in one NetLogo session becomes non-deterministic

Affected code

• src/main/providers/ModelRegistry.scala
• src/main/config/ModelRegistryLoader.scala
• src/main/LLMExtension.scala (load-config and list-models paths)

Suggested fix directions

1. Separate bundled registry state from active override state.
2. Recompute the effective registry from bundled + current-directory override each time the model directory changes.
3. Explicitly clear overrides when no override file exists in the current directory.
4. Add regression coverage for: model A with override -> model B without override -> llm:list-models must revert to bundled models.

Start by tracing ModelRegistry.loadOverride in src/main/providers/ModelRegistry.scala and how ModelRegistryLoader.scala and LLMExtension.scala handle load-config and list-models. Compare the bundled and active provider state when moving from a model directory with models-override.yaml to one without it. Done means overrides no longer leak and regression coverage verifies that llm:list-models returns bundled models for the second directory.