Before declaring RoCEv2 ready for 911 model-training bursts, what should municipal fabric engineers verify end to end?

Before declaring RoCEv2 ready for 911 model-training bursts, municipal engineers must verify end to end that lossless priorities and correct DSCP (and matching Cos) markings are in place on the AI path. RoCEv2 readiness is a fabric-and-NIC consistency problem: without lossless treatment and coherent markings, cutover merely enables a protocol that still drops or misclassifies under load. Provisioning Priority Flow Control and ECN thresholds to use DSCP 0 for the RoCEv2 AI path does not verify end-to-end lossless priorities and correct DSCP markings, because DSCP 0 is best-effort and will not receive lossless treatment, so it fails the readiness bar for drop-sensitive emergency-services trainers. Provisioning Priority Flow Control and ECN thresholds for priority 0 on AI uplinks does not verify end-to-end lossless priorities and correct DSCP markings for the RoCEv2 AI path, because priority 0 is the default best-effort lossy class and does not provide lossless treatment for RoCEv2 bursts on the civic POD. Enabling 802.3x link-level pause on AI uplinks addresses only link-level pause and does not confirm end-to-end lossless priorities and matching DSCP markings, so it does not provide lossless RDMA behavior. Exam caveat: a green NIC checkbox or lab ping does not equal production readiness—priority maps must match from GPU NIC through leaf and spine to the peer NIC. Operational check: before 911 trainer cutover, walk the RoCEv2 path with classification captures, confirm lossless Cos and DSCP consistency hop by hop, run a burst test that proves pause/ECN behavior without unexpected drops, and sign off only after those checks pass.