feat(evaluation): support scenario file references - #1696

Summary

• Add field-local top-level scenarios file refs/globs and flatten them before scenario lowering.
• Add parser, validator, schema, and generated eval-writer schema coverage for scenario refs and malformed scenario files.
• Add examples/features/scenarios with inline plus glob-loaded translation scenarios and a deterministic local CLI target.

Validation

• bun install
• Red check before implementation: bun apps/cli/src/cli.ts validate <tmp scenario eval> rejected scenarios[0] with "Scenario must be an object".
• bun --filter @agentv/core build
• bun apps/cli/src/cli.ts validate examples/features/scenarios/evals/suite.yaml -> valid 1 / invalid 0
• bun run validate:examples -> 109 valid / 0 invalid
• bun apps/cli/src/cli.ts eval run examples/features/scenarios/evals/suite.yaml --targets examples/features/scenarios/targets.yaml --target translation-cli --output /tmp/agentv-scenarios-run-LjjGCq -> PASS, 4/4 scored 100%
• Manifest inspection confirmed lowered rows: inline-portuguese-hello, french-hello, spanish-hello-world, spanish-thank-you, all status ok, score 1
• bun run lint
• bun --filter @agentv/core test test/evaluation/loaders/jsonl-parser.test.ts test/evaluation/validation/eval-validator.test.ts test/evaluation/validation/eval-file-schema.test.ts test/evaluation/validation/eval-schema-sync.test.ts -> 217 pass / 0 fail

Notes

• This stays field-local to scenarios; it does not add universal arbitrary-field file expansion.
• No providerPromptMap, scenario result grouping, artifact primitive, or Dashboard behavior changes.
• Live provider/grader dogfood was not used because the runtime change is parser/validator/example-only and the example uses deterministic local CLI output plus deterministic assertions.