feat: hosted MCP server at /mcp (gap #10)

[man4ish]

Summary

• Mounts the mcp SDK's Streamable HTTP transport at /mcp, gated by the same omni_sk_ API keys every /v1 route already trusts.
• Three tools (answer_with_citations, search_literature, list_domains) loop back to this gateway's own /v1/literature/* routes over HTTP, reusing 100% of existing rate-limit/quota/billing/BYOK logic instead of reimplementing it for a second transport.
• Deliberately does not use the SDK's OAuth-flavored auth subsystem (token_verifier/AuthSettings): that would publish .well-known/oauth-protected-resource discovery metadata implying a standards-compliant authorization server with an interactive authorization-code+PKCE flow, which omnibioai-auth doesn't actually run. Built a narrower, honest Bearer-token gate instead — see app/services/mcp_auth.py's docstring.

Bugs found and fixed while building this

1. /mcp was mounted after the catch-all router, so it was silently matched (and its own auth bypassed) by /{service}/{path} instead.
2. streamable_http_app()'s default host auto-enables DNS-rebinding protection that 404s every real (non-loopback) Host header — fixed by passing host="0.0.0.0".
3. StreamableHTTPSessionManager requires its task group to be running (session_manager.run()) before any request arrives — wired into app.main's lifespan().
4. FastAPI's root_path="/_svc/gateway" constructor override breaks Starlette's Mount child-root_path computation for everything nested under /mcp, 404ing all real (non-test) traffic even with 1-3 fixed — fixed by resetting scope["root_path"] inside MCPBearerAuthASGIMiddleware before forwarding.

Each of these is covered by a regression test (tests/test_mcp_mount_integration.py includes a full real initialize JSON-RPC handshake assertion, not just a status-code check, specifically because a weaker assertion masked bug 4 for a while during development).

Explicitly still open

A full interactive OAuth 2.1 "connector" flow (discovery + dynamic client registration + PKCE) for generic third-party MCP clients. Not buildable without fabricating capabilities omnibioai-auth doesn't have today; left for a future milestone once/if that's actually built on the auth side.

Test plan

• pytest tests/test_mcp_auth.py tests/test_mcp_server.py tests/test_mcp_mount_integration.py -q — 17 passed
• Full suite: pytest -q — 423 passed
• Manual end-to-end verification against the real app.main.app (including its real root_path) via a standalone MCP initialize handshake script

🤖 Generated with Claude Code