Release Scope and Claim Boundaries

This page is the canonical claim-scope checklist for the current development state. It keeps README, documentation, release notes, and manuscript drafts aligned with the tracked artifacts in docs/_static. If a claim is not listed here or in the referenced gate JSON, treat it as unpromoted. Here, “current” means artifact-backed and release/manuscript scoped. “Deferred” means visible as an audit or planning lane but unavailable for release notes, README highlights, abstracts, or paper conclusions until a later gate promotes it.

Current scoped claims

Claim scope for this release is intentionally artifact-limited: each release-ready claim below must be backed by the cited tracked figure, JSON report, test, or workflow gate. Open manuscript physics lanes stay visible in the guardrail artifacts, but they are not promoted by the release-readiness score.

Lane	Status	Supported claim
Linear and nonlinear benchmark atlas	release-ready for named cases	Linear growth/frequency/eigenfunction and nonlinear window statistics are validated for the tracked release cases. The nonlinear window statistics gate includes only Cyclone, Cyclone Miller, KBM, W7-X, and HSX. ETG nonlinear pilots and KAW/TEM stress lanes are not part of the release nonlinear parity claim unless a later gate index admits them.
Runtime/refactor artifact contract	release-ready as infrastructure	The large runtime and diagnostics refactor is covered as a behavior preservation claim: extracted startup, chunk, result, validation-gate, and artifact helpers keep the public runtime and NetCDF restart/append contracts stable. This does not promote new physics validation, nonlinear optimization, or performance claims.
Quasilinear diagnostics	release-ready as diagnostics	Electrostatic linear heat/particle weights, spectra, and model-selection artifacts are reproducible. The refreshed 12-case train/holdout calibration report rejects the one-constant absolute-flux family, with CTH-like and shaped-pressure external VMEC admitted only through explicit high-grid policies. Simple one-scalar saturation rules are rejected on the expanded sweep. The spectral_envelope_ridge candidate closes the declared core portfolio after excluding the Solovev and shaped-pressure stress outliers from that scoped claim: core mean relative error is about 0.280, held-out core error is about 0.275, and interval coverage is 10/10. It is retained as a scoped model-development and optimization-screening result. The full 12-case universal predictor remains unpromoted because the stress cases and rank/correlation gates do not pass. No runtime/TOML absolute-flux predictor, universal nonlinear transport model, or user-facing saturation law is promoted. Electromagnetic quasilinear field-channel normalization and KBM calibration remain future gates.
Differentiable geometry	release-ready for equal-arc parity and reduced QH/Li383 gates	The vmex -> booz_xform_jax -> GKX bridge is validated for equal-arc field-line parity where the current mboz=nboz=21 parity artifact passes QH, QI, and shaped-pressure finite-beta rows. The fixed-resolution QI row now passes after the Boozer half-mesh convention fix, with drift mismatch about 7.13e-2 against the 8e-2 tolerance, and the evaluated QI ntheta=8,16 variants pass. The shaped-pressure finite-beta eigenfrequency-gradient and quasilinear-gradient gates also pass with max relative AD/finite-difference errors about 6.4e-11 and 2.1e-4. The shaped-pressure finite-beta reduced nonlinear-window estimator gate also passes with max relative error about 2.1e-4. This is still not a broad QI transport, finite-beta nonlinear transport-gradient, or optimization claim. Reduced frequency, quasilinear, and nonlinear-window-estimator gradients pass AD/finite-difference gates on QH and Li383. The actual nonlinear finite-difference audits are startup plumbing checks with false transport-average gates; they do not validate production turbulence gradients.
VMEC/Boozer reduced objectives	release-ready for reduced gradient and UQ plumbing	The public in-memory objective path supports reduced linear frequency, electrostatic quasilinear proxy, and smooth nonlinear-window-estimator objectives through the mode-21 VMEC/Boozer bridge. The QH and Li383 holdout matrix is the citeable gate for these reduced objectives. The QL-seeded nonlinear-gradient state-control screen has admitted Rsin_mid_surface_m1 and Zcos_mid_surface_m1 as internal VMEC-state controls. The first symmetric RBC/ZBS state-to-input response fails closed, while the LASYM=true RBS/ZBC response passes as a conditioned launch mapping. This row still does not promote multi-surface/multi-alpha optimization, calibrated absolute quasilinear flux prediction, or converged nonlinear heat-flux gradients.
Stellarator optimization examples	release-ready as reduced examples plus selected optimized-equilibrium audit	The examples demonstrate differentiable reduced ITG objectives, UQ, and AD/finite-difference checks. The nonlinear objective is a reduced window-estimator path, not a nonlinear turbulence-gradient path. The selected optimized QA equilibrium now has a converged post-transient seed/timestep transport-window audit, so the production guard is closed for that scoped audit. Broad multi-surface nonlinear optimization and nonlinear turbulence gradients remain unpromoted. A VMEC-JAX transport-gradient diagnostic now also shows a measurable local boundary gradient for the aspect-6 QA restart and a solved-gate projected line-search bracket: the best accepted reduced transport metric improves by 3.55%, while the next larger step is rejected by the QS gate. This supports gate-aware projected admission. The matched long-window audit of that earlier aspect-6 accepted projected step is negative: both seed/timestep ensembles pass, but the ensemble mean heat flux changes from 9.833 to 9.891 (relative reduction -0.00585). It is therefore not a nonlinear turbulent-flux optimization claim. The companion redesign report fails closed and requires a multi-surface, multi-field-line, multi-k_y objective before another nonlinear audit can be used for promotion. A later strict top-12 QA edge audit uses that 18-point objective coverage but still fails promotion (0.58% relative reduction, uncertainty z-score 0.20). The subsequent broad max-mode-5 matrix campaign is also negative: accepted QA/ESS passes only 9/18 samples, projected weight 1e-3 fails early with 1/18 passing samples and mean reduction below 2%, and projected weight 5e-4 increases heat flux on the first completed sample. The release scope therefore remains reduced-objective and scoped-audit evidence only.
Parallelization	production-ready for independent work	Independent k_y scans, quasilinear spectra, sensitivity batches, and UQ ensembles preserve serial ordering and have solver-backed scaling artifacts. Runtime scan TOMLs may use [parallel] strategy = "batch" with axis = "ky" for this independent scan path. Whole-state nonlinear sharding is a correctness/profiler gate only.
Performance	release-ready for scoped profiler evidence	Runtime/memory panels, RHS profiler artifacts, and state-sharding identity checks are tracked. No broad nonlinear multi-GPU speedup or production domain-decomposition claim is made.

Explicitly unpromoted claims

Do not make these claims from the current artifacts:

• universal or user-facing absolute quasilinear flux prediction;

• treating refactor/test coverage as new physics validation or as a nonlinear performance claim;

• using spectral_envelope_ridge as a shipped runtime or TOML saturation option;

• electromagnetic quasilinear transport calibration for KBM;

• broad multi-surface production nonlinear heat-flux stellarator optimization; the tracked broad matrix campaign failed all selected candidate families and is recorded as negative evidence in docs/_static/broad_nonlinear_transport_matrix_negative_evidence.json;

• production nonlinear optimization without converged post-transient audits of optimized equilibria; the selected QA optimized-equilibrium audit is the current scoped exception;

• treating the historical strict QA full-sweep matched audit that stopped near t=400 as a nonlinear holdout or optimized-transport success; it remains launch-contract evidence only;

• treating the newly admitted true t=1500 growth-objective, quasilinear-objective, or nonlinear-window-objective QA candidate triplets as an optimization-success or quasilinear-calibration claim. The matched strict QA baseline now passes the same t=[1100,1500] postprocess, and all three candidate comparisons fail the 4% reduction gate: growth gives only 0.60% reduction (z=0.26), while quasilinear and nonlinear-window give -0.49% (z=-0.19) and -0.25% (z=-0.09), respectively.

• converged nonlinear transport gradients through vmex and booz_xform_jax;

• launching nonlinear-gradient campaigns directly from admitted VMEC-state controls without a separate state-to-input mapping artifact;

• treating compact nonlinear finite-difference startup audits as saturated transport averages;

• treating reduced nonlinear-window estimators or startup finite-difference audits as optimized-equilibrium nonlinear heat-flux audit bars;

• multi-surface, multi-alpha, or multi-k_y stellarator optimization from the current reduced single-fixture objective evidence;

• broad W7-X validation beyond the tracked single-flux-tube ITG windows;

• broad QI validation beyond the fixed-resolution mode-21 equal-arc parity row;

• citing even the fixed-resolution QI mode-21 row when the latest regenerated parity artifact fails, errors, or is missing;

• W7-X TEM / kinetic-electron validation;

• W7-X long-window zonal recurrence/damping closure;

• nonlinear multi-GPU speedup from whole-state sharding;

• FFT-axis nonlinear domain decomposition.

Release figure and artifact inventory

Use this inventory when deciding which figures can support release notes, README claims, or manuscript claims.

Claim family	Current release/manuscript artifacts	Boundary
Benchmark validation	benchmark_core_linear_atlas.png, benchmark_core_nonlinear_atlas.png, nonlinear_window_statistics.{png,json}, and validation_gate_index.{png,json}	Nonlinear release parity is the five-case window-statistics set only. Stress, pilot, and non-indexed example figures are not promoted.
Quasilinear diagnostics and model selection	quasilinear_*_spectrum.*, quasilinear_validated_calibration_inputs.*, quasilinear_stellarator_train_holdout.*, external_vmec_dshape_replicates/dshape_replicate_t250_ensemble_gate.*, external_vmec_circular_replicates/circular_replicate_t700_ensemble_gate.*, quasilinear_saturation_rule_sweep.*, quasilinear_candidate_uncertainty.*, and quasilinear_dataset_sufficiency.*	Electrostatic diagnostics and manuscript model selection are in scope. Runtime absolute-flux prediction and electromagnetic calibration are not.
Autodiff and differentiable geometry	autodiff_inverse_growth.*, autodiff_inverse_twomode.*, differentiable_geometry_bridge.*, vmec_boozer_parity_matrix.*, vmec_boozer_gradient_holdout_matrix.*, vmec_boozer_differentiability_claim_guard.*, nonlinear_window_fd_audit.*, and vmec_boozer_nonlinear_window_fd_audit.*	Reduced AD/finite-difference gates are in scope. Production nonlinear turbulence-gradient and broad optimized-equilibrium heat-flux claims are not; the selected QA optimized-equilibrium replicated audit is covered by the scoped stellarator-optimization row rather than by this general AD inventory row.
VMEC/Boozer objective and optimization checklist	vmec_boozer_solver_frequency_gradient_gate.*, vmec_boozer_quasilinear_gradient_gate.*, vmec_boozer_nonlinear_window_gradient_gate.*, vmec_boozer_li383_solver_frequency_gradient_gate.*, vmec_boozer_li383_quasilinear_gradient_gate.*, vmec_boozer_li383_nonlinear_window_gradient_gate.*, vmec_boozer_gradient_holdout_matrix.*, vmec_boozer_multi_point_objective_gate.*, vmec_boozer_reduced_portfolio_guard.json, vmec_boozer_aggregate_line_search_comparison.*, vmec_boozer_aggregate_alpha_holdout_gate.*, vmec_boozer_aggregate_surface_holdout_gate.*, vmec_boozer_second_equilibrium_aggregate_gate.*, vmec_boozer_aggregate_holdout_promotion_gate.json, nonlinear_gradient_ql_seed_screen.*, nonlinear_gradient_state_control_runbook.*, nonlinear_gradient_state_to_input_mapping_campaign.*, nonlinear_gradient_state_to_input_mapping_response.*, nonlinear_gradient_asymmetric_state_to_input_mapping_campaign.*, nonlinear_gradient_asymmetric_state_to_input_mapping_response.*, nonlinear_gradient_state_control_short_bracket_launch.*, nonlinear_gradient_state_control_short_bracket_launch_status.*, nonlinear_gradient_state_control_short_bracket_nonlinear_audit_status.*, vmex_transport_gradient_diagnostic.json, vmex_transport_gradient_line_search.*, nonlinear_window_ensemble_readiness_manifest.json, nonlinear_window_convergence_reports/*.json, stellarator_itg_optimization_comparison.*, and stellarator_itg_optimization_uq.*	These artifacts support reduced objective differentiability, optimizer plumbing, local UQ, explicit nonlinear ensemble-readiness blockers, and a checked state-control launch guard with a retained fail-closed symmetric negative control. They do not support calibrated saturated-flux prediction, production nonlinear turbulence gradients, direct VMEC-state launches, or optimized-equilibrium nonlinear audits beyond the selected QA candidate documented below.
Scope guardrails	benchmarks/references/gkx_1_7_release_contract.json, technical_release_status.json, release_readiness.json, and w7x_tem_extension_status.json	The frozen contract records release-scoped claims and baseline API and performance evidence. It also release-gates canonical numeric fingerprints for the promoted nonlinear, KBM branch, collision, restart, geometry, derivative, and optimization references, so textual renaming cannot hide a numerical change. It does not promote deferred physics by itself.
Performance and parallelization	runtime_memory_benchmark.*, independent_ky_scan_scaling_large.*, quasilinear_uq_ensemble_scaling_large.*, and parallelization_completion_status.*, plus nonlinear_sharding_*	Independent-work parallelization and profiler localization are in scope. Whole-state nonlinear sharding is not a production speedup claim.

Artifact-backed details

Runtime/refactor state:

• The current large refactor has extracted runtime startup, diagnostics, adaptive chunks, result assembly, validation-gate helpers, zonal-validation helpers, parallelization policy helpers, and runtime artifact boundaries into smaller tested modules. This is a maintainability and public-behavior preservation lane.

• Restartable nonlinear NetCDF append now normalizes loaded diagnostics to the persisted schema before concatenation. Monitored complex mode traces that are transient in memory and not written to *.out.nc remain absent on reload, so continuation artifacts do not mix persisted and non-persisted diagnostic fields.

• These refactor checks support release engineering only. They do not change the benchmark, quasilinear, QI, nonlinear optimization, or performance claim surface without the artifact gates listed below.

Quasilinear model-selection state:

• docs/_static/quasilinear_stellarator_train_holdout_report.json: nonlinear inputs are valid, but the one-constant absolute-flux model remains passed = false with held-out mean relative error about 6.49.

• tools/release/check_nonlinear_transport_gates.py convergence and gkx.diagnostics.transport_windows provide the reusable late-window convergence metadata required before any future holdout report can be promoted to calibrated_absolute_flux. This is a metadata/finite-window guardrail over existing traces, not a substitute for new long nonlinear simulations.

• gkx.diagnostics.transport_windows.nonlinear_window_ensemble_report provides the next guardrail for replicated windows: seed, initial-condition, timestep, or restart variants must have individually passed late-window reports and mutually consistent late means before a nonlinear turbulent-flux optimization artifact can claim robustness. tools/release/check_nonlinear_transport_gates.py ensemble is the tracked artifact wrapper for this gate.

• tools/release/check_nonlinear_transport_gates.py readiness converts tracked transport-window summaries into explicit convergence-report JSON files and a readiness manifest. The older global docs/_static/nonlinear_window_ensemble_readiness_manifest.json remains a base-window manifest, but the current D-shaped and circular case-local replicate campaigns now pass their own ensemble gates. Those case-local artifacts supersede the stale global missing-replicate message for those two cases. The QH VMEC/Boozer held-out surface/field-line campaign and the selected optimized-equilibrium audit now also pass their local seed/timestep ensemble gates.

• tools/release/check_vmec_boozer_gates.py aggregate-holdout now requires a passed replicated nonlinear-window ensemble artifact in addition to aggregate finite-difference, line-search, and held-out surface/field-line evidence before any optimized-equilibrium production nonlinear heat-flux claim can be promoted. Single-window convergence reports remain necessary but insufficient for that claim level.

• tools/release/check_nonlinear_optimization_gates.py production-guard is the explicit production nonlinear turbulent-flux optimization guard. Its tracked artifact, docs/_static/production_nonlinear_optimization_guard.json, passes release safety because reduced/startup estimators are blocked and three long post-transient replicated holdout ensembles pass: D-shaped VMEC, circular VMEC, and QH VMEC/Boozer. The selected optimized QA equilibrium contributes one accepted t=[350,700] seed/timestep replicated transport-window audit, and the strict t=1500 growth/QL/nonlinear-window candidates now close the optimized-equilibrium trace-count requirement with four qualifying ensembles. The scoped production nonlinear turbulent-flux optimization guard now promotes under its explicit 2% long-window matched-audit policy: three matched baseline-to-optimized audits pass with positive uncertainty-separated heat-flux reductions. This is scoped candidate evidence, not a broad multi-surface nonlinear transport-optimization claim.

• tools/release/check_nonlinear_optimization_gates.py gradient-evidence is the stricter nonlinear turbulence-gradient claim gate. The tracked docs/_static/nonlinear_turbulence_gradient_evidence_status.json artifact passes the replicated long-window uncertainty side but fails closed on the gradient side. The current tracked production-candidate artifact is the optimized-QA/ESS ZBS(1,0) 7.5% follow-up at t=[450,900]: all twelve runtime outputs pass, the baseline and minus replicated ensembles pass, and the central finite difference is both response-resolved (response_fraction = 0.0319) and local (fd_asymmetry_rel = 0.044). It still fails promotion because the plus ensemble spread is 0.196 > 0.15 and the propagated uncertainty is gradient_uncertainty_rel = 1.81 > 0.5. The earlier overdetermined optimized-QA/ESS RBC(1,1) 3% campaign and seed follow-up also remain failed production candidates: all runtime-output and replicated-window gates pass, but gradient_uncertainty_rel = 0.683 remains above the 0.5 gate. The companion ZBS(1,1) 3% overdetermined campaign passes uncertainty but remains nonlocal, while the overdetermined ZBS(1,0) bracket is not response-resolved. The status artifacts therefore record complete runtime coverage where expected and zero promoted controls, so this remains a failed production-candidate gate rather than a missing campaign. Until a paired post-transient artifact passes all response, asymmetry, conditioning, and propagated uncertainty gates, nonlinear turbulence-gradient evidence remains explicitly unpromoted.

• Future perturbation refreshes must use distinct artifact slugs rather than overwriting the tracked failed candidate. For example, a new coefficient or amplitude campaign should write a slug such as docs/_static/qa_ess_zbs11_rel5_nonlinear_gradient_zbs_1_1_central_fd_gradient_gate.* and a matching refreshed nonlinear_turbulence_gradient_evidence_status.json. Release prose can promote the result only if the central finite-difference artifact passes and the evidence-status JSON reports the production gradient gate as true; otherwise it remains a documented production-candidate audit.

• The tracked QA/ESS campaign is complete negative evidence, not an unfinished launch plan: all required long-window runtime outputs exist, but no boundary control passes every finite-difference gate. ZBS(1,1) is nonlocal, ZBS(1,0) is variance limited, and increasing the RBC(1,1) amplitude worsens asymmetry.

• docs/_static/quasilinear_saturation_rule_sweep.json: no simple saturation rule is accepted. On the expanded saturation sweep, the linear-weight fit is the least-bad simple rule with mean held-out relative error about 4.42; the positive-growth mixing-length rule is about 6.49 and the training-mean null is about 1.80.

• docs/_static/quasilinear_candidate_uncertainty.json: no candidate is accepted as a universal runtime absolute-flux predictor on the expanded 12-case electrostatic-compatible candidate portfolio. spectral_envelope_ridge has full-ledger mean relative error about 0.697 and interval coverage 11/12; with the declared Solovev and shaped-pressure stress outliers outside the scoped claim, its core-portfolio mean relative error is about 0.280 and held-out core error is about 0.275.

• docs/_static/quasilinear_candidate_regularization_sweep.json: the ridge-penalty sensitivity audit does not rescue that near miss. The best tested setting is now lambda = 0.5 with mean relative error about 0.689 and held-out mean about 0.764; no tested penalty is accepted as an absolute-flux predictor.

• docs/_static/quasilinear_stellarator_usefulness.json: the current stellarator-facing synthesis is scoped as scoped_model_skill_summary_not_runtime_absolute_flux_predictor. It records HSX/W7-X as admitted finite nonlinear holdouts where the simple positive-growth mixing-length rule predicts zero, records CTH-like and shaped-pressure as scoped high-grid external-VMEC admissions, keeps QA at matched-nonlinear-audit-only scope, and keeps QH excluded until grid/window convergence passes. No model-selection result is currently accepted as a universal stellarator absolute-flux predictor.

• docs/_static/quasilinear_screening_skill.json: the current correlation/ranking synthesis is scoped as screening_correlation_model_development_not_absolute_flux_promotion. It records no accepted screening model on the expanded 12-case candidate portfolio. The least-bad spectral_envelope_ridge candidate has full/held-out Spearman correlations about 0.636/0.624 and pairwise order accuracies about 0.697/0.689, below the 0.75 gates. The declared core portfolio passes the transport/coverage diagnostic but remains just below the strict rank gate. Screening skill is therefore not promoted as a runtime saturation law or universal absolute-flux predictor.

• docs/_static/quasilinear_holdout_gap_report.json: absolute-flux promotion remains explicitly blocked. The absolute_flux_promotion_requirements and screening_promotion_requirements blocks quantify the current gaps after admitting the shaped-pressure external-VMEC high-grid holdout: the absolute train/holdout mean relative error is about 6.49 against the 0.35 gate, no full-portfolio or held-out-only screening model is currently accepted, and the independent-holdout-count blocker is closed. Screening promotion still fails the rank/correlation and transport-error gates. The external-VMEC-family and non-axisymmetric external-VMEC-family coverage requirements are satisfied, but these are evidence prerequisites, not a promoted runtime absolute-flux option.

• docs/_static/external_vmec_shaped_tokamak_pressure_dt0p04_high_grid_admission_gate.json: shaped-tokamak pressure is now admitted only as a scoped high-grid holdout. The full n48/n64/n80 ladder fails because the coarse n48 trace moves the heat-flux window by about 0.469. The retained n64/n80 gates pass at t=450 and t=650; the high-grid time-horizon gate passes; and the n80 seed/timestep ensemble passes on t=[325,650] with mean heat flux about 7.16, mean-relative spread 0.0939, and combined SEM/mean 0.0463. This does not claim full n48/n64/n80 convergence and does not promote an absolute quasilinear-flux predictor.

Nonlinear benchmark state:

• docs/_static/nonlinear_window_statistics.json records five passed release-window cases. KBM and HSX use tightened gates, Cyclone Miller is tighter than the broad release envelope, while Cyclone and W7-X remain at the 0.10 release envelope pending paper-level retuning.

• docs/_static/validation_gate_index.json currently records 17 passed gate-indexed reports and 1 open report. The open report is the quasilinear model-selection status, which is intentionally not promoted to an absolute-flux predictor. The index is an audit view, not a blanket promotion of every figure under docs/_static.

• docs/_static/nonlinear_transport_time_horizon_audit.json separates long post-transient transport windows from startup finite-difference and reduced-envelope checks. Startup windows must never be described as saturated heat-flux averages.

Differentiable-geometry state:

• docs/_static/vmec_boozer_parity_matrix.json is the source of truth for the multi-equilibrium zero-beta equal-arc field-line convention gate at mboz=nboz=21. The current regenerated artifact passes QH, fixed- resolution QI, and shaped-tokamak rows. The QI row nfp3_QI_fixed_resolution_final has drift mismatch about 7.13e-2 against the 8e-2 release tolerance, and evaluated QI ntheta=8,16 robustness variants pass. The full declared QI seed campaign is still artifact-limited because three QI input variants have no bundled wout reference. The builder rejects mboz,nboz < 21 so QI is not silently evaluated on the under-resolved low-mode setting.

• docs/_static/vmec_boozer_gradient_holdout_matrix.json passes reduced linear, quasilinear, and nonlinear-window-estimator gradient gates on QH and Li383 with maximum relative mismatch about 2.7e-2.

• The VMEC/Boozer objective artifact checklist for README and manuscript use is the parity matrix, the six single-equilibrium frequency/quasilinear/reduced nonlinear-window gradient-gate figures, the combined holdout matrix, the multi-alpha aggregate objective gate, the reduced-portfolio provenance guard, the growth-vs-quasilinear line-search comparison, the positive reduced alpha-heldout and surface-heldout splits, the Li383 second-equilibrium aggregate gate, the blocked aggregate promotion JSON, nonlinear_gradient_ql_seed_screen.*, nonlinear_gradient_state_control_runbook.*, nonlinear_gradient_state_to_input_mapping_campaign.*, nonlinear_gradient_state_to_input_mapping_response.*, nonlinear_gradient_asymmetric_state_to_input_mapping_campaign.*, nonlinear_gradient_asymmetric_state_to_input_mapping_response.*, nonlinear_gradient_state_control_short_bracket_launch.*, nonlinear_gradient_state_control_short_bracket_launch_status.*, nonlinear_gradient_state_control_short_bracket_nonlinear_audit_status.*, nonlinear_gradient_state_control_bracket_sweep_status.*, optimized_equilibrium_replicates/optimized_equilibrium_replicate_t700_ensemble_gate.*, qa_no_ess_reference_replicates/qa_no_ess_reference_t700_ensemble_gate.*, qa_no_ess_to_optimized_nonlinear_audit.*, qa_ess_zbs10_rel7p5_nonlinear_gradient_zbs_1_0_central_fd_gradient_gate.*, qa_ess_zbs10_rel7p5_variance_reduction_plan.*, and the reduced stellarator ITG optimization/UQ panels. This checklist is the current boundary between objective plumbing, checked state-control launch guards, and transport prediction.

• docs/_static/vmec_boozer_reduced_portfolio_guard.json is the artifact-level guard that ties the backend-free portfolio reducer to real VMEC/Boozer rows. It requires VMEC/Boozer path/mode provenance, two field-line alpha values, two k_y samples, finite aggregate FD fields, finite growth/QL AD/FD objective gates, and an explicit non-production nonlinear claim boundary.

• docs/_static/nonlinear_window_fd_audit.json and docs/_static/vmec_boozer_nonlinear_window_fd_audit.json pass only startup finite-difference plumbing checks. Both record transport_average_gate = false.

• Finite-beta drift reconstruction, converged nonlinear turbulence gradients, broader VMEC/Boozer nonlinear transport-gradient validation, and broader optimized-equilibrium nonlinear audits beyond the selected QA candidate remain future promotion gates.

Parallelization and performance state:

• The current-branch representative linear/nonlinear refresh admits two local CPU rows and passes CPU/GPU numerical checks. Its office-GPU timing rows are explicitly blocked by measured device contention, so this refresh does not promote a new speed claim.

• docs/_static/independent_ky_scan_scaling_large.json and docs/_static/quasilinear_uq_ensemble_scaling_large.json support production independent-work parallelization for scans and ensembles.

• docs/_static/parallelization_completion_status.json is the release closure ledger for parallelization: production independent-work CPU/GPU scaling is closed, while nonlinear domain and FFT-axis decomposition remain diagnostic.

• docs/_static/nonlinear_sharding_strong_scaling_large.json is an identity and profiler-direction artifact. It shows whole-state nonlinear sharding is identity-correct but not a production speedup path for the current decomposition.

• docs/_static/nonlinear_domain_parallel_identity_gate.json and docs/_static/nonlinear_spectral_communication_identity_gate.json are diagnostic identity gates for local halo chunks and spectral split/reassemble communication layout, respectively. They are correctness prerequisites for future nonlinear domain decomposition, not runtime distributed-FFT or nonlinear speedup claims.

• docs/_static/nonlinear_sharding_profile_office_gpu.json is a tiny smoke artifact. The controlling benchmark-grid result is docs/_static/nonlinear_sharding_profile_office_gpu_benchmark_grid.json; it is slower than serial and fails final-state identity, so whole-state nonlinear sharding is blocked from production use and speedup claims.

• docs/_static/nonlinear_device_z_pencil_transport_gpu2_observable_split_profile.json is the final performance artifact for this release tranche. It passes serial-vs-sharded identity on the auto-chunked two-GPU transport-window diagnostic, but compute-only speedup remains below the promotion gate and the observable gate is dominated by scalar diagnostic overhead. This closes the current performance work as diagnostic evidence and defers production nonlinear domain decomposition to a future fused RHS/update diagnostic route.

Deferred manuscript lanes

The current manuscript/readme scope intentionally defers:

• W7-X zonal long-window recurrence/damping closure under the paper-facing initializer and observable;

• W7-X multi-flux-tube, multi-surface, and TEM / kinetic-electron validation;

• experimental W7-X fluctuation-spectrum claims through diagnostic transfer functions.

These release-scope decisions are frozen in benchmarks/references/gkx_1_7_release_contract.json; detailed W7-X extension state remains in docs/_static/w7x_tem_extension_status.json. W7-X zonal recurrence and TEM/kinetic-electron extension are explicitly deferred. The W7-X fluctuation-spectrum panel is a validated simulation diagnostic only; it is not an experimental density-spectrum validation.

Pre-release checklist

Before tagging a new public release:

1. Run the fast shard set, docs build, package build, repo hygiene, mypy, and wide coverage matrix.

2. Confirm the coverage workflow reports the package-wide 95% gate and that coverage-wide-shard-manifest.json has labeled data for every wide coverage shard.

3. Confirm README and this page agree with benchmarks/references/gkx_1_7_release_contract.json.

4. Confirm runtime/performance claims point to fresh profiler artifacts for the exact backend, device count, problem size, and identity tolerance being claimed.

5. Bump the package version before tagging; PyPI rejects duplicate versions.