Engineering feedback decision record: 3D brake-caliper study

This record turns one external engineering trial into reusable product decisions. The case used a quadratic tetrahedral mesh, a P2 displacement field, tagged pressure and clamp surfaces, and a three-dimensional linear-elastic solve. It is evidence about workflow reliability, not a request to copy case-specific code into the library.

Decisions made now

Feedback	Decision	Product consequence
a boundary carried both a marker and an imported physical tag	accepted as P0	BoundaryRegion.selection names the source of truth; imported tagged and hybrid regions use facets topologically for strong constraints and ds(tag) for weak terms
marker and tag could disagree silently	accepted as P0	BoundaryRegion.audit(strict=...) and Model.audit_boundaries(...) report facet counts, measure, midpoint bounds, integrated normal, and set differences
a pressure area required ufl.as_ufl(1) in the case	accepted	results.region_measure(on=region) is the public application-level operation
one DOLFINx XDMF grid was produced per field	accepted as P0 for serial product output	ordinary AnalysisStep.solve_result(output=...) uses AgentFEM's one-grid XDMF/HDF5 layout; U, point fields, and DG0 cell fields are Attributes of the same Uniform Grid
a CG1 result could not be written on P2 output geometry	accepted	continuous auxiliary fields are represented on the solution's nodal output grid before unified XDMF storage; unsupported higher-order discontinuous fields fail with the field name and a remediation
optional output failure erased the meaning of a successful solve	accepted as P0	output failure produces completed_with_output_errors, retains the live result, records the exception, and only raises with strict_output=True
extrema lacked a location	accepted as P1	field_extrema(..., location=True) reports coordinates, rank, global dof, sampling method, and DG0 cell identity
stress display meaning was unclear	already addressed and reinforced	default S/E/MISES fields remain explicitly documented DG0 cell-average projections without nodal extrapolation or smoothing

The intended application code is now:

pressure_surface = mesh.tagged_boundary_region(
    domain, facet_tags, tag=102, name="pressure_surface"
)
bolt_holes = mesh.tagged_boundary_region(
    domain, facet_tags, tag=101, name="bolt_holes"
)

model.pressure(16.0e6, on=pressure_surface)
model.clamp(displacement, on=bolt_holes)

boundary_evidence = model.audit_boundaries(strict=True)
pressure_area = results.region_measure(on=pressure_surface)
simulation = model.step(target=displacement).solve_result(
    output="outputs/result.xdmf",
    field_variables=("S", "E", "MISES", "SENER"),
)
peak = results.field_extrema(simulation.fields["MISES"], location=True)

This removes the need for case-owned boundary-area UFL and ordinary XDMF plumbing. Expert UFL remains available for new physics and nonstandard quantities, but it is no longer required for this standard workflow.

Deliberately not disguised as a quick fix

AgentFEM does not rename a global continuous projection as SPR or PPR. A smooth scalar MISES projection is useful for presentation, but it is not equivalent to recovering stress tensor components and then computing the invariant. Material-aware nodal recovery therefore remains a reviewed mechanics feature, with linear patch, bending, material-interface, and stress-concentration benchmarks required before it becomes a standard representation.

The same discipline applies to broader NASTRAN CTETRA10/CTRIA6 support and a mesh-study object. Both are valuable, but they require preserved PID semantics, node-order and Jacobian evidence, and convergence measures beyond a single maximum. They are not allowed to displace boundary identity and output reliability work.

Visualization and checkpoint roles

In serial, solve_result(output=...) writes one temporal Uniform Grid per frame. U and other continuous fields are point data; S, E, MISES, and other DG0 fields are cell data. The geometry is the reference configuration, so ParaView's Warp By Vector uses U once and cannot reveal a second unwarped copy hidden inside the same reader.

The low-level io.XDMFTimeSeries intentionally remains a thin DOLFINx writer. DOLFINx writes multiple Functions as separate XDMF Grids, so it is not the recommended ParaView product for a multi-field serial analysis. Existing files can be inspected with Extract Block, followed by Append Attributes, then Warp By Vector.

Under MPI, AgentFEM continues to prefer DOLFINx's collective XDMF/HDF5 path over gathering a large distributed mesh to rank zero. That path can still appear as multiple blocks in ParaView. A future collective single-grid writer must preserve mixed point/cell fields and high-order geometry. VTX/BP remains a candidate for high-order visualization, but one VTX writer requires its Functions to share an element type, so it is not by itself a complete answer for mixed P2 displacement and DG0 stress.

Checkpoint/state output and visualization output are separate contracts: checkpoints preserve continuation state and identity; visualization products optimize field discovery and presentation. Neither may silently substitute for the other.

Next evidence-bearing increments

1. Couple boundary geometry evidence with applied-load resultant, solved reaction, and relative equilibrium error in SimulationResult.
2. Define nodal_l2 as an explicitly named presentation representation, then implement tensor-first, material-aware recovery separately.
3. Add CTETRA10/CTRIA6 NASTRAN fixtures that verify PID separation, node order, positive Jacobians, area, and bounding boxes.
4. Design a mesh-study result around displacement, energy, reaction, path and regional measures, hotspot drift, and singularity warnings.
5. Add a collective visualization backend only after mixed-field and MPI ownership tests define what “one dataset” means in distributed output.
6. Extend agentfem doctor from dependency versions to the exact Python executable, imported package path, installed distribution path, and a clear warning when a source checkout shadows another installation.

Feedback intake rule

A feedback item is promoted when it supplies a reproducible case, identifies a scientific or usability consequence, and can be guarded by a stable test. It is then classified as: correctness risk, evidence gap, workflow friction, new capability, or presentation preference. Correctness and silent-failure risks come first. A useful suggestion can still be deferred when its honest implementation needs benchmarks or a larger semantic design.