Scenario validation: run probes inside a longer conversation (context depth)

[kcarnold]

Background

scripts/scenario_design/probe.ts (added in the colleague-eval merge) runs single-turn probes at depth 0 — i.e. right after the scenario's opening messages. This catches base-case behavior but not regressions that only appear once the conversation has accumulated context.

Ask

Add an optional context-depth mode to probe.ts: load an existing archetype conversation log from outputs/<scenario>_<archetype>.json (produced by simulate.ts), use it as the conversation prefix, then append the probe and judge as usual. This verifies that accumulated context doesn't cause the colleague to (e.g.) start volunteering info or drafting.

Notes

• Deferred from the colleague-eval merge (branch merge-colleague-evals) per review.
• Reuse the already-exported loadScenario / callColleague (simulate.ts) and judgeConversation (judge.ts).