fix: tolerate unknown LM Studio reasoning levels

[dgnsrekt]

Problem

rc.2 discovers no models at all against current LM Studio builds.

LMStudioCapabilitiesSchema pins reasoning values to off|on|low|medium|high. LM Studio now publishes an xhigh level:

"reasoning": {
  "allowed_options": ["off", "low", "medium", "xhigh", "on"],
  "default": "xhigh"
}

Because the enum sits inside LMStudioModelsResponseSchema, one model carrying xhigh fails safeParse for the whole payload:

models.0.capabilities.reasoning.allowed_options.3 | Invalid option: expected one of "off"|"on"|"low"|"medium"|"high"
models.0.capabilities.reasoning.default           | Invalid option: expected one of "off"|"on"|"low"|"medium"|"high"

discoverModels throws unsupported response, so the config hook logs LM Studio model discovery unavailable and contributes nothing. The provider then falls back to models.dev's built-in lmstudio catalog, offering models the user has never downloaded.

On my instance 2 of 37 records report xhigh, and discovery returned 0 models.

Fix

Accept any string for allowed_options and default.

This block is parsed but never mapped — per the v1 contract, OpenCode's reasoning flag is deliberately left unset "until the two projects publish an interoperable mapping." Validating the values strictly gains nothing today and costs total discovery every time LM Studio adds a level. The known levels stay as an exported KNOWN_REASONING_LEVELS constant for documentation.

I went with permissive rather than adding xhigh to the enum so the next new level can't reproduce this.

Verification

Regression test added — it fails on main with LMStudioAPIError: LM Studio models API returned an unsupported response and passes with the fix.

npm run validate clean (lint, typecheck, 28 tests, build).

End-to-end against LM Studio 0.4.x with 37 records, via opencode models lmstudio:

	before	after
models discovered	0 (3 models.dev phantoms)	29

Notes

Related to #34. This is a prerequisite for a follow-up PR I'd like to send adding an opt-in onlyLoaded option (#17) — that one can't be exercised at all until discovery works.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S7Wjrr4wy4Eiig1PqtHfpD