fix(skill_eval): add trigger-only mode

[mxl]

Problem

skill_eval runs every query through opencode run even after the synthetic skill has already triggered. This wastes time on downstream workflow output that is irrelevant to trigger detection, and can cause false negatives when the model later hits rate limits, timeouts, or unrelated errors.

Changes

1. plugin/lib/process.ts: Allow onStdoutChunk callback to return boolean. When true, kill the child process early (with graceful + SIGKILL fallback).

2. plugin/lib/run-eval.ts:

   • Add triggerOnly option (default true) to RunEvalOptions and runSingleQuery.
   • When triggerOnly=true and the synthetic skill name appears in stdout, immediately return true without waiting for the full process to finish.

3. plugin/lib/run-loop.ts: Thread triggerOnly through RunLoopOptions into runEval.

4. plugin/skill-creator.ts: Expose triggerOnly boolean schema arg on both skill_eval and skill_optimize_loop tools.

5. Tests: Add early-stop test in process.test.ts; add triggerOnly: true to run-loop.test.ts fixture.

Testing

• npm run build passes
• npm run test:ts: 44 pass, 0 fail
• Live eval on video-summary with GPT-5.4: 3/3 passed, triggerOnly stopped runs immediately after trigger