evaluate returns NaN in parallel for interior points the 50-neighbour walk cannot find (18 at np2, 120 at np4) — pre-existing

[lmoresi]

Surfaced by the #556 adversarial review, confirmed pre-existing (identical counts at the merge base and at the PR head, so #556 neither causes nor fixes it).

On a 3-D simplex box deformed x → x**4, 3000 interior query points: uw.function.evaluate returns 18 NaN at np=2 and 120 NaN at np=4, for points points_in_domain classifies as interior. Serial is clean.

Two contributing mechanisms, both worth their own look:

1. The RBF fallback rung cannot rescue them. vars = mesh.vars.values() (function/_function.pyx:1273) is a generator over a WeakValueDictionary, exhausted by the dofcount loop at :1280, so the fallback's for var in vars: at :1402 iterates nothing. (Point locator: reject foreign points cheaply, locate once, and containment-check the cell hint (items 1-3 of #551, fixes the #432 class) #556 fixes this line for the serial case it newly exposed; the parallel hole is wider than that one line.)
2. The 50-neighbour cap loses genuinely interior points. Measured separately during the same review: 6 of 3000 interior points on a graded 3-D mesh are not found at the evaluation tolerance, because the walk gives up at k = min(n_cells, 50). Those are exactly the points that then need the fallback.

Why it matters: a NaN from evaluate propagates silently into whatever consumes it — a projection, a swarm update, a diagnostic — and the parallel-only nature means it appears when a model is scaled up, not during development.

Related: #551 (the locator umbrella; this is a sibling of its items 1 and 5), #556 (fixes the serial exposure of mechanism 1).

Underworld development team with AI support from Claude Code

[lmoresi]

Update: this is fixed by PR #556, not separate-and-pre-existing as filed.

The dead RBF fallback rung (mechanism 1 here) turned out to be the cause of the parallel NaNs too, not just the serial exposure. With vars = list(mesh.vars.values()) the counts go from 0/18/120 at np1/2/4 to 0/0/0 on the same probe. Keeping this open until #556 merges, then closing against it.

Mechanism 2 (the 50-neighbour cap losing genuinely interior points — 6 of 3000 on a graded 3-D mesh) is real and remains: with the rung repaired those points now get an RBF value instead of a NaN, which is a much better failure mode but still a fallback rather than a location. That belongs with #551's locator work.

[lmoresi]

Fixed by PR #556, merged as 3f54218. The dead RBF fallback rung (vars = mesh.vars.values() — a generator over a WeakValueDictionary, exhausted by an earlier dofcount loop) was the cause of the parallel NaNs as well as the serial exposure that surfaced it: the counts go from 0/18/120 at np1/2/4 to 0/0/0 on the same probe.

Mechanism 2 remains, tracked under #551: the 50-neighbour cap still loses genuinely interior points (6 of 3000 on a graded 3-D mesh). With the rung repaired those points now take an RBF value instead of a NaN — a much better failure mode, but still a fallback rather than a location.