CI flake: MPI_Init dies in UCX on some Azure runners (mana_0 verbs transport)

[lmoresi]

Twice in two days (PR #526 first run, PR #527 first run) the test job died at the import smoke test, before any underworld3 code loaded:

ib_iface.c:1316 UCX ERROR mana_0: iface ... failed to create UD QP ... Operation not supported
ucp_worker.c:1415 UCX ERROR uct_iface_open(ud_verbs/mana_0:1) failed: Input/output error
MPIDI_UCX_init_worker(86): ucx function returned with failed status

UCX is trying an unsupported verbs transport on the runner's Azure MANA network interface; whether a run flakes depends on which VM type the job lands on. Rerun on a fresh runner passes (measured, #526).

Suggested immunisation: pin the transport in the workflow env for the test job, e.g.

env:
  UCX_TLS: tcp,self,sm

(or UCX_NET_DEVICES=lo for the serial import smoke alone). Filed so the reruns stop being folklore.

Underworld development team with AI support from Claude Code

[lmoresi]

Duplicate of #461 (same signature: uct_iface_open(ud_verbs/mana_0) failing MPI_Init in the import smoke test on Azure MANA-NIC runners). Consolidating there — #461 has the datapoint history. These two hits plus the two on PR #466 (2026-08-05/06) make four in a week; the fix (pin UCX_TLS=tcp,sm,self in CI — a single-node runner needs nothing else) is going up as a small PR rather than more rerolls.