Empty-rank mesh construction, tensor stats, global metric moments, swarm.points label

[lmoresi]

Four fixes from the triage backlog plus the public adversarial-review round and its responses. Closes #399, closes #400, closes #351, closes #386.

The fixes

#399 — mesh construction crashed on ranks owning zero cells. Two layers: the navigation kd-tree build passed a 1-D numpy.array([]) into the KDTree memoryview, and the KDTree constructor itself could not represent an empty cloud (&points[0][0]). Empty point sets are now first-class: (0, dim) clouds build no C++ index and answer queries honestly (found=False, infinite distances); 1-D input gets a legible error. StructuredQuadBox(elementRes=(2,2)) now constructs, integrates and solves at np3/np4 (PETSc itself is fine with empty ranks). Follow-ups from the review, also in this PR: the three sentinel sanity guards were off-by-one (> vs >= — the empty-tree sentinel masqueraded as valid index 0), and _get_domain_centroids produced NaN for starved ranks, which gather_data silently stripped — de-aligning the rank/row table and making _route_by_nearest_centroid silently migrate particles to a rank owning zero cells (demonstrated live in the review; now a huge-finite sentinel keeps the table aligned and the starved rank unreachable).

#400 — stats() crashed for every genuine tensor variable. Two stacked defects: the Frobenius loop indexed the flat component count against the structured array axis, and the temp variable was built through the enhanced factory whose __getattr__ blocks underscore delegation. The review then caught my first fix under-measuring SYM_TENSOR norms (off-diagonals must count twice) — the final form reads the structured (N, d, d) view, correct for both tensor kinds, validated against numpy.

#351 — mesh_metric_mismatch moments were partition-dependent. The equidistribution target and the returned moments are now globally reduced (median via gather-to-root); starved ranks participate in the reductions with empty contributions; non-simplex refusal is rank-symmetric via mesh.isSimplex. Known limit, documented in place: a metric containing MeshVariable data still cannot run with a starved rank — that needs the evaluate layer (#405).

#386 — deprecated swarm.points labelled metre magnitudes with mesh.units. Scaled values are now labelled meter, the mesh.X.coords convention; verified all three coordinate views agree physically under a km-scale model.

Adversarial review (public, round 1)

Posted in full below. All four primary fixes held their core claims; the review found and demonstrated live: the SYM_TENSOR under-measurement, the silent particle mis-route, the sentinel off-by-ones, a deadlock in #351's empty path with field metrics, and — most valuable — the honest map of where empty-rank support ends: construction, solve, integrals, swarm populate and stats all work; the first evaluate/estimate_dt still fails (rank-asymmetric raise in get_min/max_radius). Everything in that next layer is scoped in #405 (which also ties in #314).

Verification

• Serial battery (kdtree, stats, swarm-units, follow-metric, smoothing): 56 passed.
• np4: test_0700 (incl. new starved-rank construction test) + test_0750_global_statistics (incl. new mesh_metric_mismatch: rank-local A_target/A_mean sums make the returned δ moments partition-dependent #351 partition-independence test): 27 passed; np2: 13 passed.
• Direct repro: StructuredQuadBox(elementRes=(2,2)) constructs at np3 and np4 (crashed on development).
• Level_1 + tier_a gate after the response commit: 443 passed, 0 failures.

Reminder for the merge: manual close needed for #399, #400, #351, #386.

Underworld development team with AI support from Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[lmoresi]

Adversarial review — round 1 (public record)

Independent skeptic pass over the branch diff, instructed to REFUTE each fix; live probes at np2/np4 including starved-rank configurations.

#399 — KDTree empty clouds

• Class internals HELD: every dereference of the C++ index audited (build_index, both query paths, dealloc-on-NULL verified via live-instance count); rbf interpolators reach the index only through the guarded queries.
• Consumers BROKEN (fixed in response commit): get_closest_cells indexed _indexMap with the empty-tree sentinel — the sanity guard was off-by-one (> where >=), letting sentinel 0 pass as a valid index (demonstrated at np4: two ranks raise IndexError, two return cells). Same off-by-one at three sites; all now >=.
• Silent mis-route BROKEN (fixed in response commit): _route_by_nearest_centroid at np4 on a 2x2 box silently migrated 18 particles to a rank owning zero cells — NaN centroid from the empty mean, silently stripped by gather_data, de-aligning the rank/row table. Now a huge-finite sentinel keeps rows aligned and starved ranks unreachable (finite, not inf: infinities poison kd-tree bounding boxes).
• Honest map of what breaks NEXT on a starved rank: construction OK, Integral OK, Poisson solve OK, swarm populate OK — the wall is the first uw.function.evaluate / estimate_dt (get_min/max_radius does an empty-array reduction, rank-asymmetric raise → deadlock). Scoped as Empty-rank support, layer 2: evaluate/points_in_domain, radii/centroid reductions, gather_data NaN-stripping #405.

#351 — global metric moments

• Reductions, rms normalisation, median gather, alignment sums: HELD — all five moments bit-identical across ranks for analytic metrics.
• BROKEN (documented limit + partial fix): a metric containing MeshVariable data deadlocks with a starved rank — evaluate is collective and the empty rank must skip it (calling it instead hits the Empty-rank support, layer 2: evaluate/points_in_domain, radii/centroid reductions, gather_data NaN-stripping #405 radii failure; broken either way until that layer lands). Documented in place; lifts with Empty-rank support, layer 2: evaluate/points_in_domain, radii/centroid reductions, gather_data NaN-stripping #405.
• BROKEN (fixed in response commit): the non-simplex NotImplementedError had become rank-asymmetric (starved ranks sailed into the allreduce while populated ranks raised — hang demonstrated at np4). Now refused rank-symmetrically via mesh.isSimplex before any collective.

#400 — tensor stats

• Non-symmetric TENSOR vs numpy: HELD to 1e-15, including at np4 with starved ranks; temp lifecycle clean.
• SYM_TENSOR BROKEN (fixed in response commit): the flat read counted each Voigt component once — measured 'Frobenius' 2.536 vs true 3.232. The structured (N,d,d) read counts mirrored off-diagonals twice; regression test pins the exact value.

#386 — swarm.points label

HELD: all three coordinate views (points / coords / mesh.X.coords) agree physically under a 1000-km-scale model; the save path is confirmed migrated off self.points; no consumer depends on the old label. Two test-hygiene nits (deprecated kwarg, guarded assertion) fixed in the response commit.

Outside the diff

gather_data NaN-stripping as a general landmine, the collective-crosstalk failure mode (a caught rank-asymmetric raise pairs the survivors' pending collective with the next one the failed ranks enter — observed producing garbage with no error), and the unguarded empty-source rbf/kNN paths — all recorded in #405, which also links #314 as the same family seen from the DMInterpolation side.

Response commit

ad63bc9; the level_1+tier_a gate (443 passed) and the np2/np4 batteries were re-run after it.

Underworld development team with AI support from Claude Code

[ss2098]

Professor @lmoresi, this adversarial review is very helpful. Since the remaining failures are in the empty-rank / collective-evaluation family tracked in #405, I will not try to modify the core evaluation path directly.

One place where I may be able to help is with small validation/regression work once the fix direction is clear: running the starved-rank reproducers at np2/np4, checking macOS behaviour, and helping add or update focused tests around the empty-rank cases without touching the solver internals.

I'll keep my contributions limited to small, measurable test/API/documentation fixes unless there is a specific low-risk piece you would like me to try.