Particle stress history: refill reads the right rows after a removal (#784), and the inflow datum reaches the particles that entered (#783)

[lmoresi]

Two defects in the particle Lagrangian stress history, found on the cross-slot cell of the 2026-09-23 domain-of-applicability sweep.

#784, the refill after a removal. PETSc removes a point by copying the last point into the freed slot. Swarm.repopulate kept its coordinate array in the original order while reading the reconstruction values in storage order, so a particle created in a call that also removed took the values of unrelated particles. The Lagrangian flavour caps over-full cells every step, and any clamped outflow pile-up trips the cap, so on an open flow every refill was garbage. The proxy's per-cell least-squares fit averaged it down, which is why the closed-flow validations did not see it. Fix: replay the same moves on the local arrays.

#783, the inflow datum. The flavour declared applies_inflow_value = False; a refilled inlet cell carried a reconstruction from stale neighbours. The manager now runs the refill itself after the advection, so it knows which particles it created, and gives inflow_value to every particle whose back-trace leaves the domain where the boundary velocity crosses inward. The trace is one step for a particle that moved and one cell crossing for one created this step. A no-slip wall has no boundary velocity and a free-slip wall only a tangential one, so a trace through a wall does not count.

Maxwell Poiseuille channel, 50 steps, stress_transport="lagrangian":

	per-particle shear-stress error, median / max	inlet column, proxy error
before	0.12 / 4.8 (sign flips; field max 3.2)
#784 fixed, no datum	0.008 / 0.125	0.17
#784 and #783	0.008 / 0.125	0.002

Tests: test_0068::test_a_refill_after_a_removal_reads_the_right_neighbours (97 before, 1e-8 after) and test_0075_lagrangian_history_inflow_value (baseline 0.01 against 0.002 measured, negative control 0.1 against 0.17). The particle-history, stress-transport, restart and integration-point files pass (71 tests).

Also moved: _write_inflow (with its units reduction) and _nondim_timestep to the base class. One extra commit replaces the deprecated mesh.data in the forward flavour's geometry stamp, which failed the Charter gate on this branch.

Not fixed here, marked TODO(BUG): the particle flavours write the evaluated psi_fn into non-dimensional storage without a units reduction. Not covered: a parallel test with a rank that creates no particles.

The cross-slot cell is not rescued. Re-run with both fixes: clean to step 26, then a linear-solve divergence and a 329 s step, stalled by the guard, with the stagnation point jumping. That is the same failure family as the nodal flavour on that cell at solvent fraction zero, so it is a solver question, not a refill one. The sweep's lagrangian column ran with #784 present and is not a measurement of the scheme.

Fixes #783, fixes #784.

Underworld development team with AI support from Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_017kSkAq7oJ5J3XisuvLBovo

[lmoresi]

Adversarial review before the push, two independent passes over the diff. What they found and what we did:

Defects, fixed in the pushed commits

• The inflow datum was written into non-dimensional storage without the units reduction the integration-point flavour applies. _write_inflow moved to the base class and the particle flavour now carries _components and _psi_units; _nondim_timestep moved with it so dt is reduced too.
• An MPI rank holding no particles raised on reshape(0, -1) before the collective calls and would have deadlocked the others. Reshapes are dimension-explicit.
• The stated "no datum" baseline of 4.2 had been measured with Swarm.repopulate: after a removal the coordinate rows no longer match the value rows, so every refill in that call reads unrelated particles #784 present. Re-measured with Swarm.repopulate: after a removal the coordinate rows no longer match the value rows, so every refill in that call reads unrelated particles #784 fixed: 0.17. The test's negative control and the PR table use that number; the one-step-horizon figure quoted earlier is withdrawn.
• The created-particle count was a swarm attribute set only inside repopulate, stale whenever the control was cleared. The manager now calls repopulate itself and uses the returned count.
• The back-trace test counted every boundary. A trace along a resolved curved wall, at a corner, or where the flow separates from a wall leaves through that wall and would have stamped the inlet datum there. The boundary velocity where the trace left now decides; a wall has none, or only a tangential one.
• After a removal the fix re-located the whole swarm. It now replays PETSc's swap-with-last on the local arrays.
• The Charter deprecated-pattern gate fails on this branch on a line from 36eb788; fixed in its own commit.

Recorded, not fixed here

• The particle flavours write the evaluated psi_fn into non-dimensional storage without a units reduction (TODO(BUG) in Lagrangian.update_post_solve).
• A trace that wraps through a periodic seam reads as entered, as in the integration-point rule (Integration-point stress history blows up at the inlet of a plain channel (Wi 1 and 5, Re 100); nodal and grid histories are stable there #745) (TODO(DESIGN)).
• _cell_radii is isotropic and not refreshed on deform; on a cell elongated along the flow a particle created at its downstream edge keeps the reconstruction.
• No parallel test with a rank that creates no particles.
• The new test_0068 case lands in a tier A file unreviewed.

Checked and fine: PETSc's removal semantics (verified empirically), the descending removal order, storage order after advection, return_coords_to_bounds in place on box and facet paths, the collective sequence, the symmetric-tensor column layout, both tests failing on the old build for the stated reason.