feat: record and replay mode for reproducible runs

[JNK234]

Roadmap priority: B6

Summary

Record every prompt/response pair in a run, and replay them offline for an identical re-run without API calls.

Why it matters

Three concrete uses:

• Reproducibility — an LLM model is not currently reproducible even with a fixed random seed. Replay makes a published run exactly repeatable.
• Demos without keys — present a model at a workshop or conference with no network and no API key.
• Classroom — students explore a recorded run at zero cost.

mesa-llm has an equivalent (@record_model decorator, SimulationRecorder, plus a CLI replay viewer), and it is one of the genuinely good ideas in their codebase.

What needs to be done

1. Recording: commands like llm:start-recording <file> / llm:stop-recording, capturing prompt, response, model, provider, timestamp, and agent identity.
2. Replay: llm:load-recording <file> puts the extension into a mode where calls are served from the recording instead of the network.
3. Decide the matching rule on replay — strict call-order sequence, or content-addressed lookup (which converges with caching, B5).
4. Behaviour on a miss: error, fall through to a live call, or return a placeholder. Must be explicit.
5. File format — JSONL is the obvious fit and is diffable.
6. Interaction with clear-all and multiple agents; recordings must survive being shared alongside a .nlogox.

Open questions

• Should recording be automatic under BehaviorSpace, given B7 (experiment logging) wants much the same data?
• Does replay need to reproduce timing/latency, or return instantly? Instant is more useful; timing matters only for demo realism.