neo MCP: `run` / `evaluate` need a job handle (`{jobId}` + `poll`) for work that outlives the cap

What happened: twice, a 3–4-step awaited batch inside run came back as "Device did not respond within 60s" while the in-page loop kept running and finished correctly (results later read from a localStorage stash). The 30 s cap is by design (#2670), but the caller has no sanctioned way to learn the outcome of work the page completed after the cap.

Ask: when the body outlives the cap, return {jobId} and add poll(jobId) (or wait_job). That retires the fire-and-poll workaround and the "tool-call timeout ≠ failure" gotcha.

Environment: BrowserOS neo 0.49.5.0 · Chromium 148.0.7988.97 (arm64) · macOS · client = Claude over MCP.