adapt/relax follow-ups from the #488 re-review: empty-rank eval_metric skip (3 sites), reconnect invariant guard, vacuous partition test

[lmoresi]

Non-blocking findings from the #488 re-review (full detail in the review comment on the PR), tracked here so they survive the merge:

1. Empty-rank eval_metric collective skip — latent hang at three sites in _adapt_nested (discretisation_mesh.py: nvb-native if cur_h.size:, edge_split if centroids.shape[0]:, and the sbr path which breaks out of the level loop entirely — the worst variant, that rank leaves for good). At np>1 with a field/expression metric, eval_metric is global_evaluate (collective); a cell-less rank skips it while peers enter. Needs a rank with zero cells to fire, hence latent. Fix pattern: the Adapt-on-top: closure-free edge_split engine, reconnection repair, and interface-pinned relaxation #488 collective-emptiness treatment (allreduce the predicate; all ranks call or none do).
2. reconnect.py:781 rank-local invariant guard ("a shared point was deleted") — assert-class, fires only on internal contract violation, but would hang peers if it ever fires. Route through the allgather-verdict pattern.
3. test_pinned_set_is_partition_independent is vacuous — it allgathers the union on every rank then compares the union's size to the max over ranks of that same size; identical by construction, cannot fail. Rewrite as a genuine cross-np fingerprint (the moved-vertex test added in the Adapt-on-top: closure-free edge_split engine, reconnection repair, and interface-pinned relaxation #488 fix currently covers the defect it missed).

Underworld development team with AI support from Claude Code