feat(engine): add layer-based scenario selection (evaluation.scenarios_per_layer)

[nymeria-ai]

What

Adds evaluation.scenarios_per_layer config to suite YAML schema, allowing suite authors to control how many scenarios are randomly selected per evaluation layer.

Why

Previously, scenario selection was either "run all" or done externally by consumers (e.g. agentalent's blind shuffle + slice). This led to unbalanced evaluations — potentially 3 execution scenarios and 0 reasoning.

Suite authors should control the evaluation structure.

How

New YAML config

evaluation:
  scenarios_per_layer:
    execution: 1        # pick 1 random from execution scenarios
    reasoning: 1        # pick 1 random from reasoning scenarios
    # self-improvement: omitted → all included

Rules

• No evaluation config → ALL scenarios run (backwards compatible)
• Layer omitted from scenarios_per_layer → all scenarios for that layer included
• Layer with count → that many randomly picked (seeded PRNG for determinism)
• Order follows YAML definition order (not hardcoded)

Changes

• types.ts: New EvaluationConfig interface on SuiteDefinition
• schema.ts: Zod validation for evaluation.scenarios_per_layer
• loader.ts: New pickScenariosByLayer() function (exported)
• index.ts: Export pickScenariosByLayer + EvaluationConfigSchema
• ARCHITECTURE.md: Docs section + YAML example
• 8 new tests covering all selection behaviors

Test results

Test Files  16 passed (16)
     Tests  226 passed (226)

Breaking changes

None. The evaluation field is optional with full backwards compatibility.