fix: surface LM Studio tool capability modes

[agustif]

Summary

• advertise tool calls for every discovered LM Studio LLM because LM Studio documents native and default tool-use support
• classify model keys as native, default, or unknown from the typed native v1 capability record and include that signal in structured discovery logs
• document why the plugin does not disable default support or guess at malformed tool calls by scanning assistant text
• extend unit, live OpenCode config, packed-package, and native npm resolver checks for the tool capability contract

Contract

trained_for_tool_use: false does not mean tools are unsupported. LM Studio uses its default tool format for these models, with model-dependent reliability. OpenCode exposes only a tool_call support boolean, not native-vs-default quality, so discovered LLMs use tool_call: true and diagnostics preserve the distinction.

The plugin does not claim runtime malformed-call detection: LM Studio can return malformed tool-call text as ordinary assistant content, and the installed OpenCode plugin API exposes no unambiguous failure signal for that case.

Primary LM Studio reference: https://lmstudio.ai/docs/advanced/tool-use

Verification

• npm run release:check
• npm run smoke:opencode
• npm run smoke:opencode:acp
• packed candidate chat, ACP, and native npm resolver on OpenCode 1.17.7 and 1.17.9
• packed candidate Microsoft TUI Test: 5/5 views
• unit suite: 27/27
• coverage: 92.41% statements, 83.63% branches, 97.14% functions, 92.64% lines
• candidate SHA-256: bdd2d02df3ed881de18eb48a929618d74fd4098001b1f8a410add2990a612e05