Read one run with a bounded page of attempt results

The run's pinned identity, its per-scenario rollup, and one stably ordered page of attempts with an explicit has_more. Attempts are keyset-paginated so the query count stays constant as a suite is re-run. Also returns linkable_sessions: this AI agent's most recent Talk sessions and matched phone-test sessions that can be linked to an attempt as human evidence (newest first, at most 20, with linkable_sessions_page.has_more), carrying only the session id, kind, status, turn count and timestamps. Never returns the run's idempotency key, request fingerprint or pinned contract. Without recordings:read, attempt transcript, recording_ref and facts are omitted, assertion actual values and rubric rationale are omitted, and linked evidence omits notes. Status, verdicts and pagination remain available at the base scope, identically for API keys and OAuth MCP callers.