Dynamic Cohesive Fracture Architecture

This document fixes the first AgentFEM fracture scope before implementation is allowed to grow sideways. The target is a finite-strain, dynamically loaded solid separated along a declared weak path. It is not a generic fracture framework and it is not a phase-field implementation.

Scientific target and evidence

The first computational target is the fixed weak-interface problem of Wang, Fineberg, and Needleman (JMPS 203, 106213, 2025). The later experimental target is Wang, Shi, and Fineberg, Science 381, 415--419 (2023), DOI 10.1126/science.adg7693. Its Dryad dataset 10.5061/dryad.7wm37pvz6 contains crack-speed, SED/KED, Mach-cone, wave-speed, and material-response data. Experimental conditions used for calibration must be separated from retained prediction conditions.

The JMPS model description establishes the essential numerical problem:

• an isotropic hyperelastic sheet;
• remote impact tension;
• an edge pre-crack;
• a zero-thickness cohesive surface directly ahead of the pre-crack;
• separation traction determined by displacement jump;
• a transition controlled strongly by the cohesive length scale and the stored energy in the body.

A mesh alone cannot define this problem. Reproduction additionally requires the bulk constitutive equation, cohesive equation and parameters, thickness or two-dimensional reduction, complete loading history, damping/mass-scaling choices, stable-step policy, and the crack-tracking algorithm.

Architectural decision

The feature is split into four independently verifiable consumers:

Finite-strain bulk residual        Paired-facet cohesive residual
          |                                  |
          +---------- force assembly --------+
                             |
          lumped mass + central difference lifecycle
                             |
       accepted state / energy / output / checkpoint

The existing ExplicitDynamicsStep remains the lifecycle owner. A new finite-strain provider supplies a residual recomputed from the current deformation; a cohesive interface supplies an additional explicit internal force. Existing linear explicit dynamics keeps its present provider and Golden result. Fracture does not create a second private time loop.

Study continues to declare second-order solid dynamics and the dimensional assumption. SolutionProcedure declares explicit central difference and, later, its stability policy. Finite strain is a kinematic/formulation choice; the cohesive surface is an interface model. Neither becomes a new overloaded Study name.

State ownership

The bulk compressible Neo-Hookean model is stateless. Every cohesive integration point owns committed and trial values of at least:

• maximum effective opening;
• irreversible damage;
• dissipated interface energy;
• optional regularization state, only when explicitly requested.

An attempted update is replaceable. Only an accepted time increment commits it. A failed step, rejected stability check, or restart boundary must not partly advance interface damage. Checkpoints bind state arrays to a stable interface-pair identity, quadrature rule, law schema, mesh fingerprint, and physical time.

The first executable identity now hashes ordered reference-facet geometry, normal orientation, length, and the quadrature contract. DOLFINx block-dof numbers are recorded only as execution metadata and are no longer treated as scientific state identity. This permits a valid serial dof renumbering while rejecting a different interface or law. The identity is orientation-sensitive by design and now drives both cross-partition state and force assembly.

That identity now has a collective portable-state consumer. Facets visible on several ranks receive one deterministic owner chosen from the ranks that can assemble them. The accepted maximum-opening values are stored once per physical key with the law, quadrature contract, size, and SHA-256 evidence. The same file is exercised by a two-rank write followed by a one-rank restore with a different facet order.

The split bulk graph does not normally ghost the opposite interface trace: the duplicated sides are topologically disconnected. The MPI consumer builds one physical-node owner schedule, exchanges only the remote traces needed by locally owned facets, evaluates each sorted physical facet key on one balanced owner, and returns sparse nodal contributions to the displacement owner with numeric MPI_Alltoallv. Only owned displacement entries are written. Energy uses a scalar reduction; state remains physical-facet keyed. A two-rank Explicit run and two-rank-to-one-rank transient restart exercise this complete path, while a nonuniform compression/softening case compares every assembled node force with the serial kernel. This removes dense volume-node payloads; it does not yet claim an extreme-scale neighborhood-collective optimum.

interfaces.BilinearCohesiveLaw and interfaces.CohesiveTransaction retain the original scalar Mode-I material-point contract. The paired-facet kernel now adds one explicit vector-kinematics layer rather than inferring shear behavior from a case:

cohesive = fracture.cohesive_force(
    split,
    U,
    law,
    normal_hint=(0.0, 0.0, 1.0),
    tangential_stiffness=Kt,
)

For a scalar Mode-I law this recommended factory defaults to tangential="tie", a non-degrading penalty constraint for strict Mode-I studies. This is the conservative default because a scalar law declares no shear strength or Mode-II fracture energy from which AgentFEM could infer a shear failure law. mode_i_cohesive_force(..., tangential="free") remains the backwards-compatible frictionless-slip model. tangential="degraded" is an explicit normal-damage-driven approximation for users who intend tangential release without introducing a calibrated mixed-mode law.

The normal-driven degraded mode is intended for paths that remain close to Mode-I. Its tangential force and tangent are exact for the declared response, but the scalar fatigue transaction does not carry an independent shear damage-dissipation history. Significant cyclic shear separation should use the separate experimental mixed-mode energy-range driver, which consumes complete local jump extrema and accepts only proportional or near-proportional cycles. The permanent penalty tie is likewise a kinematic Mode-I declaration, not a general post-failure interface model.

Several named split surfaces use the same public decision path through fracture.cohesive_forces(...). Each named law independently selects the strict scalar Mode-I or mixed-mode response while the returned collection keeps commit, rollback, output, energy, and checkpoint operations atomic.

Every vector response exposes JUMP_N, JUMP_T, TRACTION_N, TRACTION_T, DAMAGE, and MODE_MIXITY. The same arrays are produced by serial and physical-facet-owner MPI consumers. audit_mode_i_kinematics(...) can reject a nominal Mode-I model when tangential jump exceeds a declared ratio.

Mixed-mode and model well-posedness

interfaces.mixed_mode_bilinear_cohesive(...) is the first P1 mixed-mode material-point and global-facet route. It uses a quadratic nominal-traction initiation criterion. After initiation, the selected mode mix is retained as state and the failure energy follows either Benzeggagh--Kenane or power-law interaction. Compression uses a separate penalty and does not advance damage; regularized Coulomb resistance is optional and declared in the law summary. An optional residual tangential fraction is a true parallel elastic penalty: its stiffness is present from the virgin state, its work remains stored, and it does not alter the declared Mode-II fracture dissipation. The monotonic path scope is proportional or mildly changing mode mix. The cyclic extension adds local cohesive GI/GII ranges and BK/power interaction for proportional peak/valley cycles and an explicit station-resolved driver for ordered closed non-proportional cycles. The latter never reconstructs an unresolved waveform from extrema.

The external structural promotion route is now explicit rather than implicit. benchmarks.DelaminationBenchmarkSpec defines DCB, ENF and MMB specimen families. DCB/ENF classical compliance derivatives are automated analytical oracles; compliance_energy_release_curve(...) recovers structure-level \(G=P^2(\mathrm dC/\mathrm da)/(2b)\), and MMB requires an independently declared Mode-I partition. Acceptance also checks cohesive-zone resolution and artificial dissipation. MixedModeBendingCurve separately reads a source- identified external trace and compares load, displacement and Mode-I fraction under required tolerances. These contracts do not claim that a published finite-element or experimental curve has already been reproduced.

Damage and solver state still use begin/commit/rollback. Mixed-mode checkpoint fields are stored by the same physical facet and quadrature identity as the Mode-I maximum-opening state, so a different rank partition does not become a new material history. Checkpoint schema v5 also records the tangential mode and stiffness. Older checkpoints predate vector kinematics and restore directly only into the historical free-slip consumer. upgrades.migrate_cohesive_checkpoint(...) can produce an auditable v5 free, tie, or degraded state from physical-keyed v3/v4 data; changing shear physics requires an explicit acknowledgement and records a source digest and migration provenance. It rejects mixed-mode migration because initiation mixity cannot be reconstructed from scalar opening history. Legacy index-keyed v1/v2 states must first be restored on their original free-slip model and re-saved with physical facet identity.

Interface splitting can create several disconnected bulk bodies. Before a solve, audit_split_interface_rigid_modes(...) constructs each body's exact translation/rotation basis and measures the rank of declared Dirichlet and intact interface constraints. A free middle body therefore fails preflight instead of being hidden by a singular linear solve. The remedy must be a physical boundary or shear-carrying interface, not an arbitrary point fix.

For quasi-static post-peak paths, FiniteStrainCohesiveArcLength augments the existing bulk-plus-interface residual with one physical load unknown and a spherical constraint. It reuses the analytical interface tangent, strong-boundary elimination, rollback, and checkpoint state; force control and arc length are distinct procedures rather than switches embedded in a material law.

Interface topology

Ordinary continuous displacement degrees of freedom cannot represent a jump on a shared facet. The MVP therefore uses two geometrically coincident sides with independent displacement degrees of freedom. The mesh adapter must:

1. split the declared weak path and duplicate its nodes before constructing the DOLFINx mesh;
2. retain source-side and source-facet identities;
3. pair negative and positive facets by explicit identity, using geometry only as a checked reconstruction aid;
4. establish a deterministic normal and node/quadrature permutation;
5. reject missing, duplicate, ambiguous, or noncoincident partners;
6. assign every physical pair one balanced deterministic MPI owner.

For a conforming two- or three-dimensional mesh, the path can be derived from a declared cell partition with split_conforming_cell_interface(...). Shared edges or triangular faces are recovered and checked automatically. Abaqus ELSET plus element-face SURFACE definitions and Gmsh volume/interface physical groups now lower directly to this neutral split-interface contract. The first rigorous imported 3D route is tetrahedral: the external face set is used as an independent check that the declared surface exactly equals the cell partition boundary before coincident nodes are duplicated.

Because the split sides are topological exterior facets, a normal DOLFINx interior-facet dS form is not by itself the global cohesive element. The current 2D line and 3D triangular-surface consumers assemble paired-facet nodal forces explicitly and add them to the same residual vector as the bulk force. This also makes equal and opposite interface force and energy tests direct. A later mixed-dimensional or custom-kernel implementation may replace the assembler without changing the public CohesiveSurface contract.

Bulk dynamics

The bulk route is Total-Lagrangian hyperelasticity, currently including Neo-Hookean and Mooney--Rivlin energies:

\[ \rho_0 \ddot{u}_i = P_{iJ,J} + \rho_0 b_i, \qquad P = \frac{\partial\psi}{\partial F}. \]

At every time increment, bulk internal force is assembled from the active displacement. The recoverable bulk energy is the integral of the same potential used by that residual. The current quadratic u.T K u / 2 diagnostic is valid only for the linear provider and must not be reused for finite-strain fracture.

Initial acceleration is computed from equilibrium of the transferred state, not silently set to zero. Prescribed displacement, velocity, and acceleration histories must remain kinematically compatible. The stable increment records separate estimates from body elements and cohesive stiffness, their safety factor, and the controlling entity.

Near incompressibility and two-dimensional reduction

The existing mixed u-p Neo-Hookean static formulation is not automatically an explicit formulation: pressure has no ordinary inertial equation and the current mixed solve is monolithic. It must not be attached to central difference merely because it avoids locking statically.

The original V4 mechanism Golden uses a local finite-strain Neo-Hookean plane-stress reduction at nu=0.49. At every material point a positive thickness stretch is solved such that P33=0; the in-plane stress, energy, and Schur-condensed acoustic tangent derive from the same condition. The same benchmark can now consume an injected incompressible plane-stress Mooney--Rivlin material using the Wang--Fineberg--Needleman Eq. (17) energy, without copying its mesh, cohesive, time-integration, energy, or crack-observer workflow. Neither reduction by itself proves freedom from in-plane locking on arbitrary meshes. A trusted broader route still requires an F-bar/selective- volumetric treatment or a deliberately designed stabilized explicit u-p scheme. Artificially lowering bulk modulus to obtain the desired crack speed remains forbidden.

The current 2D Neo-Hookean implementation is plane strain. A thin-sheet claim requires either:

• a local thickness-stretch solution enforcing the declared finite-strain plane-stress condition, with its energy and consistent in-plane response; or
• a thin three-dimensional solid model.

The first option is now implemented and locally verified. The first affine thin-3D cross-check is also executable: a true three-dimensional Q1 cuboid at two thickness/cell layouts recovers the condensed nominal stress and energy to near machine precision, including transverse zero traction. This validates the homogeneous reduction, not a three-dimensional crack front, bending, out-of-plane instability, or general near-incompressible locking behavior.

Preload and step transition

Preload and dynamics are ordinary ordered analysis steps connected by a typed state-transfer object, not a special preload_then_dynamic solver. The transfer records:

• source and destination step identities;
• displacement and mesh/configuration convention;
• transferred constitutive/interface state;
• initialized velocity and acceleration;
• force imbalance and energy before/after transfer.

The first acceptance test is a held prestrained body: after transfer with no release or impact, it must remain in equilibrium without an artificial wave. A fully dynamic smooth preload is retained as an independent cross-check.

Wave speeds

Reference unstretched isotropic speeds are useful V1 checks but are not a complete supershear classifier. In a prestrained finite-deformation state, incremental bulk speeds are direction-dependent and must be obtained from the instantaneous material tangent/acoustic tensor using a declared reference or current-configuration convention.

The Rayleigh speed of a prestrained, incrementally anisotropic half-space is not obtained reliably by inserting one effective modulus into the classical isotropic formula. fracture.principal_surface_wave_speed() now solves the reference Total-Lagrangian Stroh decay problem for a principal, traction-free 2D half-space. It recovers the classical isotropic Rayleigh speed and removes the repeated-attenuation-root factor before accepting a secular root. It also rejects a base surface carrying nominal traction. This distinction matters: the V4 cohesive interface carries normal preload before release, so the traction-free oracle is evidence for the wave implementation but is not silently substituted as that interface's prestrained crack-speed boundary.

Energy ledger

Every saved history distinguishes:

• bulk recoverable strain energy;
• kinetic energy;
• cohesive recoverable energy;
• cohesive irreversible dissipation;
• external work, including prescribed-motion work;
• declared numerical damping or stabilization dissipation;
• residual balance and normalized balance error.

The first declared damping perturbation is mass proportional, f_d = alpha M v_mid. It is disabled by default. Its accepted nonnegative work is reported separately as numerical_damping_dissipation and cannot disappear inside a generic balance residual.

No unlabelled total_energy is sufficient. Interface dissipation is integrated from the same traction--separation law that supplies forces. At complete bilinear separation its value must equal the declared fracture energy times reference interface area.

Crack observations

Crack position is an observation derived from the ordered interface state, not a new solver unknown. The first definition uses a declared damage/opening threshold and spatial interpolation between cohesive points. Speed is fitted over a physical time/length window; a single newly failed element is never a reported instantaneous velocity. Every history records threshold, window, mesh spacing, and uncertainty/smoothing policy.

sub-Rayleigh, supershear, and spall-like are classifications with evidence. Spall-like separation additionally considers the fraction and time span of near-simultaneous interface failure; it is not merely a very large finite-difference crack speed.

Verification gates

V0 -- local mathematics

• rigid rotation objectivity of bulk stress/energy;
• finite-strain patch tests;
• uniaxial/biaxial Neo-Hookean paths;
• exact cohesive envelope area;
• loading/unloading/reloading and compression closure;
• cohesive commit/rollback/restart identity;
• pre-crack and intact-interface initialization.

V1 -- waves

• unstretched longitudinal and shear waves;
• directional incremental waves after homogeneous prestrain;
• mesh, element, and time-step convergence;
• numerical versus acoustic-tensor speed, with an initial target near 2%.

V2 -- energy

• no-fracture conservation;
• one-interface separation balance;
• convergence of the complete energy ledger.

V3 -- classical fracture guardrail

• sub-Rayleigh Mode-I propagation;
• no numerical crossing caused by an unstable increment;
• stability across mesh, time-step, and declared damping sweeps.

V4/V5 -- publications

• JMPS parameter trends and separation-mode transitions first;
• Science Dryad field and history comparisons second;
• calibration cases and retained prediction cases remain distinct.

The V5 exchange product is a sealed DynamicFractureEvidenceBundle: accepted interface trace, wave-speed convention, complete energy columns, comparison records, and copied field or report artifacts. Its provenance seal establishes byte integrity, not scientific acceptance. Publication/FEM coordinate registration is an explicit AffineCoordinateMap; rectilinear comparisons reject unit/configuration mismatch and preserve domain masks.

Until V0--V3 pass, the feature is experimental. validated is reserved for the named benchmark and parameter range whose evidence is published.

First 30-hour foundation

The weekend foundation is intentionally narrower than a supershear claim:

1. freeze this architecture and a machine-readable capability scope;
2. implement and test one irreversible Mode-I cohesive law and transaction;
3. implement deterministic 2D paired-interface topology and its validation;
4. assemble equal-and-opposite interface nodal forces for one segment and a strip of segments;
5. connect current Total-Lagrangian Neo-Hookean internal force to the existing explicit lifecycle without changing linear Explicit;
6. replace quadratic finite-strain energy with the constitutive integral and establish the typed fracture energy ledger;
7. add a held-prestrain transfer test and a one-interface dynamic opening test;
8. add checkpoint identity for cohesive state;
9. define V1/V3 benchmark specifications and record missing author inputs;
10. only then run the first sub-Rayleigh-to-supershear parameter exploration.

Implemented in the first foundation slice

• an irreversible bilinear Mode-I law with exact envelope area, unloading, reloading, compression closure, precrack initialization, and material-point commit/rollback/restart;
• an array-level conforming-interface splitter that duplicates only the declared positive side and retains the original-to-duplicate identity;
• deterministic coincident line-facet pairing with explicit orientation and rejection of shared, missing, or ambiguous node identities;
• a two-point 2D interface kernel with equal-and-opposite force and integrated recoverable/dissipated energy;
• serial and MPI-reference DOLFINx dof adapters and one composite bulk-plus-interface residual;
• an executable split-mesh adapter for triangular and quadrilateral cells, with automatic recovery of independent coincident displacement dofs from retained DOLFINx input-node identity;
• a Neo-Hookean Total-Lagrangian Explicit provider that recomputes internal force from the active deformation and retains the existing linear Explicit provider unchanged;
• a visible body/interface stable-increment screening estimate;
• typed bulk/kinetic/cohesive energy channels plus trapezoidal natural-load work, strong prescribed-motion work, and absolute/relative balance error;
• amplitude-driven Explicit constraints that impose displacement, midpoint velocity, whole-step velocity, and acceleration from one declared history;
• cohesive auxiliary state in the shared transient checkpoint envelope, including coincident input-node and physical-facet keys for rank-count changes;
• quasi-static displacement to Explicit u/v/a transfer with an equilibrium guard or an explicitly declared release mode;
• threshold-interpolated crack position, window-fitted crack speed, Mach-angle helper, and a separation classifier that requires independent spall evidence.
• the analytical Neo-Hookean material tangent and homogeneous small-on-large acoustic tensor, including explicit pull-back/push-forward of reference and current propagation directions. This supplies the V1 analytical oracle used by the numerical arrival-time benchmark below; it does not substitute for a prestrained Rayleigh secular solution.

This slice is still experimental. The MPI dof adapter now uses a cached physical-owner schedule with sparse trace/force payloads, and conforming paths can be recovered from cell partitions. Direct Abaqus/Gmsh surface lowering, extreme-scale neighborhood-collective profiling, MPC/contact/weak-constraint work, general near-incompressibility, and full thin-3D fracture remain gates.

DistributedDofMappedCohesiveForce.performance_profile() makes the current parallel workload reviewable without turning noisy CI wall time into a pass/fail criterion. It reports per-rank owned facets, required and remote trace nodes, numeric payload counts, and facet imbalance. PETSc -log_view supplies timing for the named constitutive, vector-assembly, and matrix-assembly events. The combination separates a decomposition or communication regression from normal runner timing variation.

Automated V1--V3 guardrails

The first named verification ladder is now executable:

• V1: a longitudinal wave packet is propagated by finite-strain Explicit at zero and ten-percent held prestrain. The measured reference-coordinate speed converges toward the material acoustic-tensor prediction and the 80-cell errors are below two percent;
• V2: no-fracture wave energy improves under spatial refinement. A smooth prescribed separation then drives one interface to complete failure; its dissipation equals the declared Gamma and the final balance error converges below 1e-5;
• V3: a precracked long interface advances multiple facets. Its threshold-interpolated, seven-frame fitted speed remains below 0.8 c_R under 40-to-60-facet refinement, explicit time-step refinement, and a small declared mass-proportional damping perturbation. Damping work is a typed energy channel and final balance errors remain below 5e-4.

This remains V3 for the named compressible plane-strain cohesive strip, not a general dynamic-fracture validation. Supershear claims are evaluated by the separate V4 mechanism gate below.

Executable V4 mechanism gate

benchmarks.jmps_weak_interface_transition_v4() now runs three cases through one public finite-strain Explicit lifecycle:

1. a homogeneous 12% plane-stress preload followed by precrack release gives a contiguous crack-like front at approximately 0.96 c_R;
2. the same body and interface with a zero-slope smooth remote impact gives a resolved front at approximately 1.10 c_s and 0.56 c_d, while no more than 4% of the ligament fails in one increment;
3. a weaker, larger-cohesive-length interface under the same impact fails across the ligament within one thickness shear-wave time and is classified spall_like, rather than assigning physical meaning to its super-dilatational apparent front speed.

Every case transfers preload without an energy jump and closes the complete energy ledger within 0.5% at the declared 30-facet mechanism resolution. The crack-speed fit window spans at least three interface cells at the prestrained shear-wave speed. The classifier also rejects an apparent speed above c_d as unresolved_discrete_failure unless independent distributed-spall evidence is present.

This is an experimental V4 numerical mechanism benchmark. It establishes the crack-like/supershear/spall distinction in the intended architecture; it does not claim curve-level reproduction of Wang, Fineberg, and Needleman's unpublished input deck. Promotion to publication reproduction still requires the authors' dimensions, exact cohesive law and parameters, impact history, mesh sequence, and post-processing convention, followed by mesh/time-step convergence and a thin-3D check. The immutable evidence and tolerances live in knowledge/benchmarks/jmps_weak_interface_transition_v4.json.

Two-dimensional V4 convergence contract

benchmarks.jmps_weak_interface_convergence_v4() is the opt-in refinement contract beyond the inexpensive two-cell-thick mechanism ladder. It uses a 30x10 near-isotropic baseline, 40x14 and 60x20 near-proportional spatial refinements, and a 30x10 run with the explicit increment halved. A lower smooth impact is used because the screening impact becomes distributed spall when transverse wave propagation is resolved.

All four runs preserve a resolved supershear regime, fewer than ten percent of the intact facets fail in one increment, and final complete-ledger energy errors remain below 0.034%. The peak of a moving local-speed window is retained for regime classification because it detects whether the front crosses the prestrained shear-wave speed. It is not used as the convergence observable: that maximum is sensitive to initiation and short transients. Refinement instead compares a least-squares propagation speed over the same reference- coordinate interval, x=[0.8, 1.8], in all four runs. The fits have R^2=0.990--0.997; the two successive spatial changes are 1.00% and 2.41%, and halving the explicit increment changes the representative speed by 0.093%. Both the declared 10% spatial and 2% temporal gates therefore pass, so the benchmark reports mechanism_preserved=True, speed_converged=True, and accepted=True. This establishes the declared two-dimensional numerical mechanism and observation contract. A cohesive-zone resolution study, thin-three-dimensional publication geometry, author parameters, and external curves remain necessary before claiming publication-level reproduction.

The transient lifecycle advances path-dependent external work at every accepted increment but assembles bulk and kinetic energy only at retained history frames. It also reuses the accepted explicit residual when evaluating prescribed-motion work. On the fixed 30x6 profiling case this reduces bulk residual assemblies from 345 to 173 while preserving the speed, failure, and energy observables bit for bit; local wall time fell from about 31.84 s to 16.07 s. Rank-local timing evidence is exposed in the Step and benchmark summaries. The profile also shows that cohesive force assembly is below 0.1% of wall time; finite-strain UFL/PETSc bulk residual assembly is the next performance target.

The next performance work is therefore:

1. retain the residual-reuse count as a per-commit regression;
2. investigate compiled-form/vector reuse around the bulk residual without changing constitutive or commit/rollback semantics;
3. keep the short construction smoke per commit and run the roughly 6.2-minute full refinement contract in scheduled or release-gate CI;
4. retain the collective workload profile and PETSc event names on larger interfaces, then evaluate a distributed-graph neighborhood collective without changing its physical identity, force, energy, restart, or benchmark contracts.

V5 public-data evidence boundary

V5 is now represented by reusable software contracts rather than a premature validated=True flag:

• Dryad version 235603 is pinned by DOI, CC0 license, 26 file names, sizes, SHA-256 identities, and scientific roles;
• a dependency-free XLSX reader inventories stored values and cached formula results without making pandas part of the FEM runtime;
• the V4 case can retain accepted facet-center opening, traction, damage, and dissipated-energy-density histories in a portable NPZ trace;
• front ensembles combine damage, opening, and dissipation thresholds and expose observer spread;
• curve, Mach-angle, and rectilinear-field comparisons share RMSE, NRMSE, correlation, overlap, and interpolation metadata;
• standard result semantics now include nodal V/A and cell KED, alongside finite-strain SENER.

The executable data audit is examples/science_supershear_v5_protocol.py; the human protocol is docs/research/science_supershear_v5.md; and the exact machine handoff is knowledge/research_tasks/science_supershear_v5.json. The research analyst, not the solver, must reconstruct spreadsheet units, freeze calibration and prediction conditions, choose scientifically justified tolerances, and attribute disagreement.

The explicit lifecycle also accepts a live derived-field group and refreshes SENER/KED/J or other supported finite-strain cell fields immediately before each saved frame. This closes the reproducibility and measurement plumbing needed to begin V5. It does not close the scientific gates. Remaining software-side blockers are full thin-three-dimensional fracture and general near-incompressibility validation, direct Abaqus/Gmsh internal-surface ingestion, extreme-scale neighborhood-collective profiling, and improved V4 spatial speed convergence. Affine publication registration, cell-partition interface recovery, sparse physical-keyed MPI force/state, and cross-rank-count cohesive restart now have executable contracts.

Inputs to request from the authors

Priority order:

1. complete computational input/source and postprocessing scripts;
2. exact bulk and cohesive equations with every parameter;
3. dimensions, thickness, pre-crack, and weak-interface construction;
4. preload/impact histories and the applied boundary convention;
5. plane-stress, plane-strain, shell, or 3D assumption;
6. element family, mesh sequence, and smallest interface spacing;
7. crack position/speed and field-processing definitions;
8. damping, mass scaling, stable-time-step factor, and boundary treatment.