feat: BYOK provider routing, token metering, and the model catalog

[man4ish]

Summary

Closes design audit gap #4 in full, as three paired PRs across omnibioai-auth (key-reveal endpoint), omnibioai-rag (actual Claude/OpenAI client routing), and this repo (request validation, the reveal-and-forward step, token-event emission, and the model catalog).

• build_rag_query now accepts model="claude"/"openai" -- only together with use_own_key: true, since there is no platform-wide key for either provider, only an organization's own BYOK key. Either alone is rejected with a clear field pointing at whichever one is missing.
• _handle_billable_literature_call reveals the organization's decrypted key from omnibioai-auth's new internal endpoint immediately before forwarding a BYOK-routed call -- resolved before the concurrency slot and quota are touched, so a request that can't get a key never consumes either. The key is attached to the one RAG request body only; never stored, logged, or visible in the headers forwarded downstream.
• A BYOK-routed call that gets real token counts back from RAG also emits llm.tokens.input/llm.tokens.output usage events (V1Store.emit_usage gained quantity/unit params for this, defaulting to the unchanged 1/"requests") -- never fabricated for the default path, where these stay absent exactly as before.
• model_source/token counts in the public /v1/literature/answers response now relay whatever RAG actually reports, instead of the old hardcoded "omnibioai_gpu"/None/None.
• GET /v1/models lists claude/openai alongside default, price: null for all three -- BYOK means an org pays its own provider directly; this platform still charges nothing on top (same "don't invent a dollar figure" rule as everywhere else this gap has touched pricing).

Test plan

• pytest tests/test_literature_contract.py -- 57 passed, including the full BYOK validation matrix (model without use_own_key, use_own_key without a BYOK model, both together accepted and passed through, default path's RAG body has no model key at all)
• pytest tests/test_v1_literature.py -- 78 passed (71 existing + 7 new: reveal call + RAG forward with the key attached, 503/400/502 error mapping for reveal failures, token-usage event emission alongside the answer event, confirming the default path never emits token events, and confirming /search still ignores model/use_own_key entirely)
• Full repo suite -- 406 passed

🤖 Generated with Claude Code