Add scenario validation pipeline and make systemPrompts reviewable

[kcarnold]

Problem: System prompts for colleague chat scenarios were stored as single long strings with \n escapes, making them hard to review and diff. More importantly, there was no systematic way to validate that a scenario's system prompt actually produces the intended colleague behavior across different participant types (e.g., does it resist drafting requests? stay in character with a confused newbie? avoid info-dumping on a passive ack?).

Changes:

1. systemPrompt → systemPromptLines (scenarios.json) Store prompts as arrays of strings for readability. studyConfig.ts joins them at load time so the runtime ScenarioConfig type and all consumers (chat route, evalColleague) are unchanged.

2. Scenario design pipeline (scripts/scenario_design/) Four manually-sequenced scripts for validating scenario prompts:

   • generate.ts: situation description + criteria → scenario JSON entry
   • simulate.ts: runs 4 participant archetypes (eager, lazy, confused, boundary-pusher) through multi-turn conversations with the colleague
   • judge.ts: evaluates conversations against 8 criteria using generateObject + Zod for structured verdicts
   • fix.ts: analyzes failures and proposes minimal prompt edits

   criteria.md defines scenario-agnostic behavioral rules (information
   gating, refusal to draft, fact consistency, tone, etc.) intended to
   double as paper appendix material.

   All scripts use the existing ai SDK + openai provider — no new deps.
   Run with: npx tsx scripts/scenario_design/<script>.ts

3. Cross-pollinated the two scenario systemPrompts to align on:

   • UNPOLISHED tone for both colleagues
   • No promises / "I'll get back to you" for both
   • Patience with new users for both
   • User role descriptions matching taskInstructions