refactor(eval): invokeDataset + per-type example classes

[jariy17]

What & why

Refactors the monolithic core.eval.simulate into two composable EvalClient calls the handler orchestrates, and models each dataset type as a class that owns its own parse + run + ground truth.

Stacked on aws#2000 (feat/eval-batch-simulate) — base that branch, so this diff is only the refactor.

• core.eval.invokeDataset replays a dataset (invoke per example, one session each) and returns { sessions, invoked, failed } with neutral inline ground truth.
• The handler composes invokeDataset → startBatchEvaluation. No Core method calls a sibling.
• Per-type example classes (PredefinedExample / SimulatedExample) implement an Example interface; DatasetLoader classifies each row by shape and news the right class. Dispatch is a switch, not a map.

CLI command structure (unchanged for users)

agentcore eval batch-evaluation simulate \
  --runtime-id <id> --payload-template '{"prompt":"{input}"}' \
  --dataset <path.jsonl | dataset-id> [--dataset-version <v>] \
  --evaluator <id...> --name <name> \
  [--qualifier PROD] [--header 'x-a: 1'] [--bearer-token -] [--user-id u] \
  [--description ...] [--kms-key-arn ...]

Output: { batchEvaluationId, status, examplesInvoked, examplesFailed }.

What changed

Area	Change
src/core/eval/dataset/	new — types.ts (Example, RunContext), predefined.ts, simulated.ts, load.ts (DatasetLoader), run.ts (runExamples)
src/core/eval/simulate.ts	deleted — split into the above; Scenario→Example, runScenarios→runExamples
src/core/eval.tsx	simulate method → invokeDataset + private readDatasetText
handlers/eval/types.tsx	SimulateInput/Result → InvokeDatasetInput/Result on CoreEvalClient
handlers/eval/.../simulate/index.tsx	composes invokeDataset + startBatchEvaluation
testing/TestCoreClient.tsx	setSimulateResponse/simulate → setInvokeDatasetResponse/invokeDataset

Design decisions

1. Dispatch is a build() switch, not a registry map. A new DatasetSchemaType member makes the switch non-exhaustive → build lacks an ending return → compile error (TS2366). No assertNever, no satisfies Record.
2. readDatasetText reuses the existing read path — readLocalDatasetFile (already used by updateDatasetExamples) for local files, downloadDatasetToTemp for dataset ids — and drops the combined loadDatasetFile.
3. Ground truth stays neutral — invokeDataset returns InlineGroundTruth; the handler wraps it into SessionMetadataShape for batch. A future ondemand simulate handler adapts the same output to EvaluationReferenceInput.
4. Simulated is scoped out — SimulatedExample's constructor throws NotImplementedError at load, so the user gets a clear "not replayable yet" instead of a per-row "has no turns" misdiagnosis.

Deliberately not in this PR (follow-up): folding invokeRuntime/normalizeRuntimeInvokeRequest into a RuntimeInvoker static class. Kept the existing free functions to keep the blast radius on the dataset seam.

Bug fixes carried from the simulate work

• Sparse multi-turn ground truth kept in position — one GroundTruthTurn per turn carrying input.prompt, so a 3-turn example whose only expectation is on turn 3 is graded against turn 3, not turn 1. (The old .filter(...) renumbered turns.)
• Both-turns-and-actor_profile row is refused (AWS's SDK silently drops the actor profile).
• Empty ground truth omits the union member (no { inline: {} }).

Testing

$ bun run typecheck        # clean
$ bunx oxlint src/core/eval/dataset   # clean
$ SIMULATE_INGESTION_WAIT_MS=0 bun test
  1401 pass, 0 fail, 2 snapshots

New tests:

• src/core/eval/dataset/dataset.test.ts — loader classification (predefined/simulated/both/neither/dup-id/missing-id/bad-JSON/empty/blank-lines), constructor validation, turn-replay order, sparse-turn position, no-GT, and a golden snapshot of the full inline ground-truth shape ([golden]).
• simulate.test.tsx — asserts the invokeDataset → startBatchEvaluation composition: runtime-level flags reach invokeDataset (no evaluator/name leak), sessions + wrapped ground truth reach startBatchEvaluation, and "nothing invoked" refuses to grade.

Mock-free: example classes are tested by constructing them and passing a fake invokeOnce.