feat: Support inline eval definitions

[Copilot]

Experiments can now reference evals defined outside the repository’s generated eval registry. Inline evals include a name, project-local directory, and optional config/test path overrides that resolve from the CLI working directory.

• Experiment config

  • Allows evals entries to be either built-in eval IDs or inline eval objects.
  • Adds shared eval config types for prompt metadata.

• Eval resolution

  • Resolves built-in eval IDs through the existing registry.
  • Resolves inline eval paths relative to process.cwd().
  • Supports inline config, configPath, and testPath.
  • Preserves sandbox spoofing behavior by normalizing inline evals into the same runtime shape as generated evals.

• Validation + docs

  • Adds focused coverage for built-in lookup, cwd-relative inline paths, and custom config/test paths.
  • Documents inline eval usage.

export const experiment: ExperimentConfig = {
  name: 'Local project experiment',
  description: 'Run an eval from the current project',
  models: ['gpt-5.5'],
  evals: [
    {
      name: 'local-button-eval',
      path: './evals/button',
      config: {
        prompt: 'Update the local project to use a Primer button',
      },
      testPath: 'button.eval.test.ts',
    },
  ],
  treatments: [],
}