[justin-carper]

Summary

Fixes #119. When a model can't take a thinking level, the provider now sends it as the param the model actually reads. Example: thinking: medium on grok-4.6 was sent unchanged; Cursor ignored it and ran at its default (high). The provider now sends effort: medium.

Behavior

• resolveControls takes the model's advertised params. If thinking isn't a valid value of the model's thinking param, or the model has none, it moves to the first reasoning param (effort/reasoning) that accepts the value. xhigh maps to extra-high where Cursor spells it that way.
• An explicit value on any effort/reasoning param wins; thinking is dropped.
• A value nothing accepts is left in place, and the warning lists the accepted values.
• If the model's params aren't known (no discovered catalog), nothing changes.
• The language model reads the model's params once per instance from the latest discovered catalog (falling back to the built-in list), and logs each adjustment once via warnOnce.
• The delegate and cloud tools are unchanged; their thinking arg is already documented as true/false only.

Verification

• npm run prepublishOnly: 656 tests pass, typecheck + build OK.
• New tests: 9 in test/controls.test.ts, plus 1 in test/language-model.test.ts that seeds a catalog and asserts grok-4.6 receives effort: medium.
• Mutation checks: removing the remap, the catalog wiring, or the any-param explicit check each fails a named test.
• Not run: a live Cursor turn. The premise that Cursor ignores params a model doesn't advertise comes from the grok 4.6 thinking/reasoning level #119 report.

Closes #119