feat: add capability-based benchmarks

[Copilot]

Benchmarks group curated scenarios by capability, exposing actionable performance breakdowns across models instead of a single aggregate score.

Changelog

New

• defineBenchmark API and benchmark discovery utilities
• CLI support via --benchmark and --benchmarks
• Capability-level and per-model summaries
• Benchmark metadata in JSON output

export const benchmark = defineBenchmark({
  name: 'UI interactions',
  description: 'Evaluate common UI tasks',
  models: [{name: 'gpt-5.5', reasoningEfforts: ['high']}],
  capabilities: [
    {
      name: 'Edits forms',
      description: 'Creates and updates accessible forms',
      scenarios: ['creates-sign-up-form', 'adds-form-validation'],
    },
  ],
})

Changed

• Result reporting now associates resolved scenario IDs with benchmark capabilities.
• Package exports and documentation cover benchmark authoring and execution.

Removed

• None.

[joshblack]

@copilot add support for this in the UI. Let's use our baseline experiment as a "Design System" benchmark