M1: add /v1/literature/search, /literature/domains, /v1/usage, /v1/models

[man4ish]

Summary

Part of M1, building on M0's contract-freeze/outbox work. Adds the four remaining endpoints from the design doc's public API table ("most public endpoints are missing"):

• POST /v1/literature/search — billable ("1 search", ~1/10th of an answer per the design's pricing table), retrieval-only. Forwards to the companion omnibioai-rag PR's new mode="search" (no LLM call at all). Shares the exact same idempotency/quota/usage lifecycle as /literature/answers via a new _handle_billable_literature_call helper, factored out so the two billable endpoints can't silently drift apart on that sequencing. Bills literature.search, not literature.answer.
• GET /literature/domains — the same omnibioai-rag study list /literature/studies already exposes (left untouched for backward compatibility), reshaped under the design doc's public name ("domain"). Free, rate-limited, never billed.
• GET /v1/usage — the gateway's first synchronous call into omnibioai-billing, forwarding to its existing GET .../subscription/usage-limits and translating included/used/remaining per resource. Free. Gated on usage.read — already registered in omnibioai-auth's Permission Registry as "reserved, not yet enforced by any route" (the same state dataset.read/model.use/workflow.execute were in before this gateway's own IAM Foundation integration); this is its first real consumer. Estimated-charge-in-dollars is deliberately omitted — would need billing's cost-summary period math, not yet done, and reporting it wrong would be worse than omitting it.
• GET /v1/models — a static one-entry catalog (no upstream call) for the one model path that actually exists today. price is null, not a placeholder number — the design doc's pricing section says to measure real GPU cost first (M0, 1,000 representative questions), which hasn't been run.

Known follow-up: usage.read is newly wired into SERVICE_PERMISSION_MAP as policy-engine context, matching the established pattern, but whether any role is actually granted usage.read today is an omnibioai-policy-engine configuration question outside this repo — flagging so it isn't missed before /v1/usage is expected to work end-to-end for real callers.

Depends on: the companion omnibioai-rag PR (mode="search") for /v1/literature/search to return real results — this gateway change is forward-compatible either way (it just forwards mode: "search" and translates whatever comes back).

Test plan

• pytest — 353 passed, including a pre-existing test (test_pr13_role_tier_permission_forwarding.py) that parametrizes over every SERVICE_MAP key and would have caught a missing SERVICE_PERMISSION_MAP entry for billing
• ruff check . — clean
• New coverage: contract-translation tests for build_rag_search_query/build_public_search, and route-level tests for all four new endpoints (auth, rate-limit, billing lifecycle, upstream-error mapping, free-vs-billed)

🤖 Generated with Claude Code