LLM Multi-Provider

StackSwap calls multiple LLM providers (OpenAI and Claude) behind a common internal surface for recommendation narrative and the guided advisor. Requests use structured outputs (JSON schema) so the response plugs directly into the report pipeline without brittle string parsing. Every call logs token usage to the report's audit columns, so cost per report is observable — not a black box.

The LLM layer powers the `justification` field on every transformation plan row and the ask-your-report chat. It runs behind server-side API routes so API keys never ship to the client, and behind a provider abstraction so we can swap or A/B providers per call type. Deterministic fallbacks exist for every LLM output — if a call fails, the report still renders with the rules engine's baseline justification.

Multi-provider + structured output + per-report token logs means StackSwap can move between OpenAI and Claude for cost or quality reasons without rewriting the recommendation pipeline. The architecture treats LLMs as a swappable input, not a core dependency.