feat(engine): scenario pool / random sampling

[nymeria-ai]

Summary

Adds a scenario pool mechanism to suite YAML that enables random task selection at runtime.

Why

Evaluation suites with fixed tests are gameable — agents (or their owners) can memorize answers. Pools let suite authors define a bank of scenarios and have the engine randomly select N per run.

Use case: agentalent.ai agent verification — 25+ creative challenges in a pool, each run picks 3-5 random tasks. Agents can't pre-script answers.

YAML Syntax

scenarios:
  # Regular scenarios work unchanged
  - id: fixed-task
    name: Always runs
    layer: execution
    input:
      prompt: Do this
    kpis: [...]

  # NEW: Pool — engine picks `count` random scenarios
  - pool:
      id: creative-challenges
      count: 3
      seed: 42  # optional: deterministic selection
      scenarios:
        - id: roast-yourself
          name: Self Roast
          layer: execution
          input:
            prompt: Roast yourself
          kpis: [...]
        # ... more scenarios in pool

Changes

• types.ts — ScenarioPool interface, ScenarioEntry union type
• schema.ts — Zod validation for pool syntax
• loader.ts — resolvePools() with seeded PRNG (mulberry32), Fisher-Yates shuffle
• index.ts — exports
• pool.test.ts — 14 new tests

Design

• Pool resolution in Loader → Runner receives flat ScenarioDefinition[] (zero Runner changes)
• Seeded PRNG for reproducible runs (no external deps)
• Count > pool size → clamp with warning
• Backward compatible — existing suites work unchanged
• All 208 engine tests pass (14 new + 194 existing)