Hand-written discrimination proof over the 227-case corpus: baseline 94/227, a correct auto-margin/over-constraint candidate 112/227 and admitted, two seeded-wrong candidates 0/227 each and rejected for regressing every previously-passing case. A real-model run against the same seam and gate is recorded too: the endpoint answers GET /v1/models but never returned a chat completion within three retried 600-second timeouts, including a minimal request sent outside this project's own client -- a completion-level stall, not an unreachable endpoint.