feat(evals): run agent scenarios through the DAGExecutor #8422

Problem

The eval suite drives the provider streaming tool loop directly. It does not
exercise agent-block input wiring, variable resolution, the executor's
retry/fallback policy, or the run/error handling above the loop. A harness change
there can regress without this suite noticing.

Proposal

Add an executor-level harness that builds a serialized workflow with an Agent
block, invokes DAGExecutor.execute, and scores the same checks.

• Location: apps/sim/evals/agent-tool-use/executor-harness.ts
• Real Start → Agent workflow; only executeProviderRequest is mocked at the
  provider boundary, so the Agent block handler, variable resolution, and the
  executor run/error handling are real.
• Reuse @sim/testing factories (createSerializedWorkflow) and the
  @sim/testing mocks already used by
  executor/handlers/agent/agent-handler.test.ts.
• Reuse the existing scoring/report so executor and loop results land in one
  report.

Acceptance criteria

• At least one scripted scenario runs through DAGExecutor end to end
• Asserts the tool calls and the Agent block's final output
• Results appear in the existing JSON/Markdown report
• Documented how to add an executor-level scenario

Follow-up

The provider boundary is mocked, so the executor's retry/fallback policy is not
asserted yet. Add a scenario whose first provider call rejects, with a block
retry config, to cover the retry path.