Agent config roundup: 'gpt-5.5' duplicated in six places, no retry/timeout on LLM calls, mock fallback doesn't cover quota exhaustion

[manjula25]

Summary

Agent configuration/resilience items, batched together.

Items

1. Model name duplicated as string literals in six places: MODEL = 'gpt-5.5' is independently declared in all four agent files (riskAgent.ts, careGapAgent.ts, sdohAgent.ts, actionPlannerAgent.ts), and scripts/eval.ts hardcodes 'gpt-5.5' twice (lines 556, 719) even though routes/analysis.ts:15 imports MODEL. Changing the model in one agent leaves the other three on the old model, and computeCostUsd returns null for unknown models — costs render as "—" while agents actually run on a different model.
2. No retry or timeout on the LLM call (riskAgent.ts:284-289 and the other three agents): a single client.responses.create; a transient 429/500 fails the whole run.
3. Fallback doesn't match its own documentation: the streamMockRisk doc comment (riskAgent.ts:204-206) claims the fallback activates when "quota is exhausted", but the code (line 277) only checks process.env.OPENAI_API_KEY presence — a quota-exhausted key throws rather than falling back.