feat(evals): record and replay live model transcripts

[sudoKrishna]

Problem

Live model runs are nondeterministic and need an API key, so they never run in
CI. Scripted runs are hand-written from scratch and drift from what a real model
actually emits. Neither gives a durable, key-free regression test for real model
behavior: today's live finding evaporates on the next run.

Proposal

Record a live run once, replay it forever, through the real tool loop.

• EVAL_RECORD=1 captures each model turn's streamed chunks per scenario to
  apps/sim/evals/agent-tool-use/fixtures/<scenario>.json.
• A replay completion feeds the recorded chunks back through
  createOpenAICompatStreamingToolLoopStream — no provider key, no network.
• Replay runs in the normal bun run test; live mode records the fixtures.
• Fixtures are committed and reviewed like snapshots: a diff shows a behavior
  change, and the eval scores the real transcript instead of a hand-written one.

Acceptance criteria

• EVAL_RECORD=1 DEEPSEEK_API_KEY=... bun run test:evals:live writes fixtures
• Replay runs with no key and no network, and is collected by bun run test
• A recorded scenario reproduces the live pass/fail it was captured from
• Fixtures are committed and reviewed like snapshots
• README documents recording, replaying, and refreshing a fixture

Follow-up

Once fixtures exist, the context/memory and subagent suites can replay real
multi-turn transcripts instead of scripting every turn.