feat(bot): unified topic processor replacing router+interpreter stack

Replace 3-layer architecture (gpt-5-nano router + gpt-4o-mini interpreter) with
single unified topic processor (gpt-4o-mini) for simplified message handling.

New components:
- HouseholdContextCache: TTL-based caching (5 min) for household config data
- TopicProcessor: Unified classification + parsing with structured JSON output

Key changes:
- Renamed ASSISTANT_ROUTER_MODEL → TOPIC_PROCESSOR_MODEL
- Added TOPIC_PROCESSOR_TIMEOUT_MS (default 10s)
- Refactored save() → saveWithInterpretation() for pre-parsed interpretations
- Removed deprecated createOpenAiTopicMessageRouter and ~300 lines legacy code
- Fixed typing indicator to only start when needed (purchase routes)
- Fixed amount formatting: convert minor units to major for rawText

Routes: silent, chat_reply, purchase, purchase_clarification, payment,
payment_clarification, topic_helper, dismiss_workflow

All 212 bot tests pass. Typecheck, lint, format, build clean.

docs/runbooks/iac-terraform.md

@@ -57,7 +57,7 @@ Keep bot runtime config that is not secret in your `*.tfvars` file:
 - `bot_mini_app_allowed_origins`
 - optional `bot_purchase_parser_model`
 - optional `bot_assistant_model`
-- optional `bot_assistant_router_model`
+- optional `bot_topic_processor_model`
 
 Set `bot_mini_app_allowed_origins` to the exact mini app origins you expect in each environment.
 Do not rely on permissive origin reflection in production.