feat(evals): add agent context eval suite #8427

Problem

The agent decides what the model sees from conversation memory, but nothing
evaluates that end to end. Unit tests cover the windowing functions
(selectConversationTokenWindow, selectConversationContextGroups); nothing
covers whether the assembled provider request actually contains prior history,
in order, with the system prompt preserved, scoped to the right conversation. A
regression in message assembly is invisible.

Proposal

Add an agent-context eval suite that drives the real Agent block through the
DAGExecutor with conversation memory on, stubs the memory read per conversation
id, and asserts the provider request.

• Location: apps/sim/evals/agent-context/
• Reuse the executor harness: a case is an executor case plus agent.memory.
• Checks: all history reached the provider, history precedes the new user prompt,
  the system prompt survived, and the configured conversation id was used.

Acceptance criteria

• 2+ context scenarios run through the real executor
• Asserts the assembled provider request (history, order, system prompt,
  conversation id)
• A wrong conversation id fails the run (verified locally)
• Results appear in a report and the README documents the suite

Follow-up

Windowing inside the memory service (sliding window size/tokens) and the
retrieval tool are the next layer; this suite covers the assembly the model sees.