Child of #695 (plan Phase 3). After gate G2 names the winning region: create district bxl1-test-cloud, rent exactly one GPU machine there, day or hourly billed. Build the full cloud body image: Windows Server, pinned GPU driver, MttVDD, Sunshine, hydrabody, WireGuard, WDAC policy. Enroll as a pet in nodes.yaml.
Run: the stack validation testbook again on this machine; hydramark (#513) at N = 1, 2, 3, 4 slots for stream_capacity; the #511 audio checks on Server (a fail caps slots at 1, does not fail the pilot); mercator-talks V0.9 end to end to a sanctioned test head (cheeky-cactus class, never production) over tunnel-direct with an in-session B1 probe.
Gate G3: stack works, stream_capacity >= 1 inside the 4.7 gate, and the interaction check passes: 10 marked taps register correctly plus a 5 s drag plus all scene transitions; tap-to-photon <= 150 ms p90 over 20 taps and within 20 ms of a same-day on-prem bxl1-test body run, measured with a 240 fps slow-motion camera; audio starts within 2 s and holds sync over 5 minutes.
No monthly commitment before this gate. Related: #504, #511, #513.