fix: prevent duplicate LM Studio models with case-insensitive deduplication

[daniel-lxs]

Summary

This PR fixes the issue where LM Studio models appeared twice in the Provider Configuration Profile when they were both downloaded and loaded. This implementation addresses the feedback from @mechanicmuthu about LM Studio's JIT Model Loading feature.

Problem

The previous implementation fetched models from two different LM Studio APIs:

• listDownloadedModels() - showed all models on disk
• listLoaded() - showed models currently in memory

This caused duplicates when a model was both downloaded and loaded, with inconsistent keys (model.path vs model.modelKey).

Solution

This PR implements a case-insensitive deduplication approach that:

• Keeps both API calls to support LM Studio's JIT Model Loading (available since v0.3.5)
• Deduplicates models by checking if any existing model key contains the loaded model's identifier (case-insensitive)
• Prefers loaded model data when duplicates are found, as it provides more accurate runtime information (contextLength)
• Maintains compatibility with JIT loading where models can be automatically loaded on-demand

Benefits

• ✅ Eliminates duplicates with smart deduplication
• ✅ Supports JIT Model Loading feature (models can be auto-loaded when requested)
• ✅ Shows all available models (both downloaded and loaded)
• ✅ Provides accurate runtime metadata when models are loaded
• ✅ Handles auto-unload after 60-minute timeout gracefully

Testing

• All existing tests pass
• Added new test case for deduplication logic
• Verified that duplicates are properly handled

Context

This replaces PR #7141 which only showed loaded models and didn't align with LM Studio's JIT loading capabilities.

Fixes #6954

---

Important

Implements case-insensitive deduplication in getLMStudioModels() to prevent duplicate model entries, preferring loaded models for accurate runtime data.

• Behavior:
  • Implements case-insensitive deduplication in getLMStudioModels() in lmstudio.ts to prevent duplicate model entries.
  • Prefers loaded model data over downloaded when duplicates are found.
  • Ensures compatibility with JIT Model Loading by keeping both API calls.
• Testing:
  • Adds test cases in lmstudio.test.ts for deduplication logic, including path-based matching and distinct name handling.
  • Verifies that duplicates are properly handled and runtime metadata is accurate.
• Misc:
  • Handles auto-unload after 60-minute timeout gracefully.

^{This description was created by for 6aba8e5. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Thank you for your contribution! I've reviewed the changes and found some issues that need attention. While the approach to support JIT Model Loading is good, the deduplication logic has some critical flaws that could cause incorrect model matching.

[roomote]

Good test case! However, could we add more edge cases to ensure robustness? For example:

• Models with similar names ("llama-3" vs "llama-3.1")
• Multiple loaded models with various duplicate scenarios
• Case sensitivity edge cases
• Models with special characters in their paths/keys

[daniel-lxs]

Hey @mechanicmuthu, what do you think of this approach?

[mechanicmuthu]

Hey @daniel-lxs

Hey @mechanicmuthu, what do you think of this approach?

Autoloading of an unloaded model is working and tested now by me.

However, the duplicate issues is there. This is due to different namings here when the model is loaded.

The bottom 2 with simpler names are from currently loaded models, the top 2 are from all models. All 4 work correctly though.

[fbuechler]

@daniel-lxs I think the solution requires more complex matching logic because LM Studio seems to adapt the model keys to make them unique if you have different quantizations of the same model downloaded. In that case the key isn't necessarily a substring of the path any more.

Example:
unsloth/Qwen3-4B-Instruct-2507-GGUF/Qwen3-4B-Instruct-2507-UD-Q3_K_XL.gguf -> qwen3-4b-instruct-2507@q3_k_xl
unsloth/Qwen3-4B-Instruct-2507-GGUF/Qwen3-4B-Instruct-2507-UD-Q4_K_XL.gguf -> qwen3-4b-instruct-2507@q4_k_xl

[daniel-lxs]

Then I think it would be great to create a properly scoped issue because lm-studio is making it quite hard for us to deduplicate the models.

@mechanicmuthu
@fbuechler

[fbuechler]

@daniel-lxs I'm sorry if I'm missing a point here but I don't think it needs to be that hard. In my opinion, the initial PR (#6956) provided a simple working solution that consistently used the model's path to prevent duplicate entries and replaced the downloaded model with more details if a loaded model with the same path was found.