Conversation auth guard

he GET /conversations endpoint accepts project_id, user_id_1, and user_id_2 as query params but never verifies the caller is one of those users. Anyone who can reach the endpoint can create or look up conversations between arbitrary users. Adding the same verify_token check used by the WebSocket upgrade (extract token from a header or query param, verify the caller's user_id matches one of the two) would close this.

Locate the GET /conversations endpoint and the WebSocket upgrade path, then compare how both use verify_token and obtain the caller identity. Add coverage showing that arbitrary user pairs are rejected while an authenticated participant can access the conversation; run the existing test suite to verify the guard.

Written by the indexing model from the issue text.