fix: support current LM Studio model metadata

[themizzi]

Summary

• accept current LM Studio native /api/v1/models records, including xhigh reasoning and newer fields
• validate LLM records separately from embedding/non-LLM records
• preserve unloaded and loaded context discovery
• include detailed fetch and validation reasons in discovery failures
• add regression coverage using current MLX/GGUF, loaded-instance, reasoning, and embedding payloads

Root cause

The release candidate schema validated every record as an LLM and restricted reasoning options to off, on, low, medium, and high. The live response included an embedding record without LLM fields and a model using xhigh, causing whole-response validation to fail and auto-detection to silently fall back.

Verification

• npm run typecheck
• npm run lint
• npm run test:run (30 tests passed)
• npm run build
• live smoke test against http://127.0.0.1:1234: 3 LLMs discovered, 2 embeddings skipped

[themizzi]

Superseded by upstream PR agustif#58