MANTA implementation, acceptance and benchmark plan

This plan accompanies the architecture and scientific contracts. The stage table specifies the full implementation target; the progress descriptions and manual identify qualified increments. The reference-test inventory is not evidence that every proposed backend already passes it. Reproducible benchmark reports live under benchmarks/. No numerical source default changes in this plan.

1. Ordered implementation increments

Stage	Concrete deliverables	Exit evidence / permitted next step
P0: shared interfaces and references	Add provider capabilities, lossless shell/image/layout adapter, source/receipt schema and a libint tile adapter. Connect a diagnostic consumer to the existing molecular JKBuilder seam. Record fixed operator ownership.	Identical libint integrals/J/K through old and new boundaries, no duplicate normalization or changed thresholds; unsupported combinations rejected before setup. Snapshot lifetime and preflight budget gates pass. No new production default.
P1a: molecular primitive kernels	Implement FP64 Boys and MD recurrence, S/T/V, cross-basis S, moments, 2c/3c/4c Coulomb for s/p. Direct J/K and a small capped reference materializer.	Analytic/high-precision and libint parity, arbitrary-density contractions and explicit memory/work receipts. Selectable experimental manta on the admitted subset.
P1b: full molecular kernel envelope	Extend pure/Cartesian scalar orbital l through 6, auxiliary l through 7, general/partial-general contractions, batched multiple targets, erf/erfc and FT; optimize only after parity.	All admitted classes tested, contraction schemes agree, full/erf/erfc identities hold, zero-screening and all permutations work. Unsupported classes use declared fallback or refusal.
P1c: method and RI adapters	Add native JKBuilder integration, spin slots, direct and streamed RI J/K, metric admission, output/citations and explicit backend selection.	RHF/UHF/RKS/UKS parity and immutable solve capture, method energy components and terminal-density checks. ROHF/ROKS and response are individually admitted, not inferred from closed-shell tests.
P2a: molecular BUPO-J	Extent envelopes, moments/translations, deterministic hierarchy, disjoint near/far traversal, paired contraction, error accounting. Near field can use either provider.	Compare arbitrary-density J to the same unscreened/direct source; isolate tails, multipole, screening and summation errors. No exchange acceleration claim.
P2b: molecular RI-BUPO-J	Paired forward/adjoint coupling, streamed metric solves, fit/rank controls, incremental builds with full confirmations.	Exact RI versus approximate-pass comparisons, adjoint/energy-derivative tests and chain-drift safeguards. Measure actual crossover versus optimized RI and Split-J.
P3a: periodic primitive substitution	MANTA shell/FT tiles inside current Gamma and whole-mesh erfc/reciprocal/GDF sources. Keep BIPOLE domains and exchange seams.	Same-source S/T/V/J/K and energy-component parity on general complex trial densities, shifted/non-TRIM meshes and distorted cells. No molecular-limit shortcut.
P3b: periodic BUPO-J and RI-BUPO-J	Bounded finite Ewald lattice M2L coefficients are qualified; production disjoint traversal, error controls, q=0/charge ledger and q-aware RI remain.	Independent Gaussian derivative quadrature through order 12, near-set transfers, split/trace/gauge and periodic Gaussian energy checks qualify the coefficient primitive. Diffuse/skew-cell convergence and whole-mesh responses remain method gates.
P3c: low-dimensional sources	Explicit slab/wire Green functions and open-direction boundary policies with dimensional stress conventions.	Same-dimensional independent references; 3D vacuum extrapolation is supplementary, not the defining reference. Keep unsupported combinations refused.
P4: CCM and chi adapters	Individual adapters for union/WSSC, neutral Bloch, neutral real-Gamma and chi four-center/RI/RIJCOSX. Prove weight factorization before BUPO aggregation; use streamed direct terms otherwise. Inventory model-specific SECCM electrostatics.	Preserve finite domains/weights and experimental gates. Same-H Fourier equivalence only where applicable. No global supercell AO matrices or four-index tensors in production. Conditional capability decisions documented per source.
P5: qualified symmetry reduction	Method-owned admission over shell orbits, stabilizers, bubble/fit/support policies; transport in generation and digestion; explicit identity-only path.	Independent unreduced same-source comparisons, arbitrary-density equivariance, failed-parent/rank/phase negative tests, measured executed-work reduction. Existing diagnostic success alone does not enable reduction.
P6a: derivatives	Nuclear gradients for molecular exact/RI then BUPO, periodic and CCM sources; stress as a distinct extension. Include Pulay, metric/projector, moment/translation, domain, gauge and seam terms.	Distorted-geometry finite differences, sum rules, cell strain and topology/rank crossing diagnostics for each selected approximation. Nonvariational sources retain their restrictions.
P6b: further acceleration	Specialized recurrence DAGs, GPU batches, distributed target tiles, optional mixed precision and measured kernel dispatch.	Same scientific gates and end-to-end error budget; host/device transfer and replicated memory included. No imported ORCA/LibintX/EXESS speedup claims or unreviewed dependencies.

P0-P2 can land without waiting for a universal periodic Fock refactor. The periodic raw short-range fitting increment reuses the existing GDF image selector with MANTA 2c/3c kernels. Analytic Boys double sums, same-source libint metric/coupling blocks and general complex fitted-response trials check Gamma and non-TRIM transfers. Signed long contractions, pure/Cartesian angular shells, independent high-l radial/Fourier integrals, image gauges, reciprocal relabeling, engaged envelope screening and simultaneous exact/ one-short resource caps retain their separate acceptance. The output owns only raw SR components; complete RSGDF assembly and method selection remain required. Its report is benchmarks/manta_periodic_fitting.md.

The subsequent complete Coulomb source assembles SR and tiled MANTA Fourier blocks with the existing jellium kernel. Independent analytic finite sums and converged bare-Coulomb reciprocal references validate Gamma and non-TRIM transfers, two numerical splits, finite-zero subtraction independent of list membership, and complex trial responses. Polynomial quadrature checks the Cartesian and high-angular Fourier assembly. Gauges, left-handed cells, engaged SR screening and joint caps remain explicit. Fitted SCF is qualified separately below. The complete-source report is benchmarks/manta_periodic_gdf.md.

The immutable tile source now reconstructs that dense source bitwise while admitting only one metric or coupling shell tile. Its fixed-projector q-channel RI consumer is checked against independent Gaussian/Boys/Fourier sums and libint blocks, including signed angular contractions, arbitrary complex indefinite spin densities, nonuniform weights, image gauges, conjugate transfers and weighted energy derivatives. Converged numerical splits agree with an independent bare-Coulomb fit. Redundant, truncated and zero-rank auxiliary spaces, significant negative-metric refusal, owned snapshots, source identities and all five exact/one-short resource caps have separate tests. Count-only memory witnesses admit streaming below the size of one full coupling tensor; they do not execute the large numerical workload. The report is benchmarks/manta_periodic_ri_q.md.

The whole-mesh fitted Fock consumer now passes independent Gaussian/libint checks on every matrix and energy component, weighted energy derivatives, mesh reordering, and full/truncated-projector image transport through the shared symmetry density-push/operator-pull interface. Opposite transfers must preserve the actual finite reciprocal support and conjugate metric, projector and inverse before the consumer is admitted. Arbitrary trial spin densities remain independent; no spatial or time-reversal reduction is used.

Fixed independent J/K selection additionally passes separate Gaussian/libint matrix and component-energy references, weighted energy variations and reciprocity for full, truncated and zero-rank fits. Gamma/non-TRIM meshes, nonuniform/zero weights, exact component decomposition, all five joint caps and one-short refusal are tested. J-only owns no exchange-transfer metric; its actual metric/coupling calls and complete admitted bound are smaller on the multik witness. Independent image transport and owned exports cross the landed shared Fock API with explicit source flags and no symmetry admission. Eighteen prechange joint J/K states preserve matrices, scalar receipts, projectors and admission counts bitwise. Timing and modeled-memory evidence is in benchmarks/manta_periodic_ri_targets.md. Trial-density diagnostics also compose each RI target with the complementary direct reciprocal source and check independent responses, energies and derivatives with both transfer conventions preserved. Mixed HF acceptance is described below. Fitted KS, physical range fits and BIPOLE parity still require their own acceptance.

Physical finite-source RI UHF now reuses S/T/V, nuclear Ewald, the fixed-spin bounded solver and the method-owned exchange correction. Independent libint raw fits, NumPy contractions and SciPy SCF check Gamma/two-point/shifted H and Li states, unequal/empty spins, the restricted limit, all energy components, and every capped epoch. AO and auxiliary image transports check complex physical snapshots. All five aggregate caps include every retained projector, metric construction once, repeated coupling passes, scalar history and terminal confirmation. Selected scientific output adds admitted per-transfer evidence without changing the caller’s budget.

An offline compatibility probe also admitted 24 actual RI UHF states (72 spin/k channels) into the then-inspected FockBuild draft. Evaluated density, method energy, fixed source, S/X frame, physical action, metric populations, immutable storage and solver-parent identities were preserved. Capped states keep their distinct next orbital projector. That probe does not land or duplicate the draft API, replace method convergence criteria, qualify BIPOLE/CRYSTAL parity, or authorize symmetry reduction. The report is benchmarks/manta_periodic_ri_uhf.md. The RI UHF cutoff factory now passes independent integer enumeration of skew, odd/even and shifted reciprocal spheres, exact-boundary/empty-source checks, pair-complete AO image membership and independent AO/auxiliary gauge transport. Explicit-source and generated runs have identical D/F and scientific source identities. Independent libint/NumPy/SciPy checks cover Gamma and shifted H/Li SCF with both exchange corrections. All five exact/one-short caps include generation, auxiliary sources, projectors and repeated solver work. Its geometric receipt remains distinct from the fitted physical source. The report is benchmarks/manta_periodic_ri_domains.md.

Public RI UHF job selection now preserves the same independently checked finite source, every fitting control and metric projector, explicit J/K choices, evaluated spin history and lossless QVF density. Joint source/output admission, dry runs, output delivery failures, truncated/zero-rank fits, independent AO/auxiliary images and capped states have separate checks. Repository tests now check the 24 public states against the landed FockBuild API, preserving method energy, physical action, source/density/frame separation and solver-parent identities without adding dense snapshots to the bounded native solve. The report is benchmarks/manta_periodic_ri_job.md.

Restricted RI HF now reuses the same immutable fitting/physical source while running the bounded RHF numerical leaf. Independent libint metric/coupling and one-electron/nuclear sources, NumPy contractions and SciPy restricted SCF cover He/Be, Gamma, full/shifted meshes and both exchange corrections. Capped recurrences, negative/empty/truncated auxiliary spaces, redundant AO spaces, equal-spin embedding, exact/one-short native/domain/output caps, independent AO/auxiliary images and nuclear-frame QVF have explicit checks. Shared FockBuild crossings preserve the restricted total density and retained S/X frame; no source symmetry certificate is inferred. Singlet UHF has the same fitted source but its own solver/storage and method identity. Sixteen existing UHF capped/converged states remain bitwise unchanged after sharing RI HF composition. Native and public benchmark costs are recorded separately in benchmarks/manta_periodic_ri_rhf.md. Fitted KS, physical range-separated fitting and the remaining capability matrix stay open.

Mixed exact/RI RHF and UHF now pass independent libint raw-fitting, Gaussian Fourier quadrature and NumPy/SciPy SCF references for both component choices, Gamma/full/shifted meshes, empty/unequal spins and both exchange corrections. The 152-case mixed suite also checks every capped recurrence, all five native/generated/public resource boundaries, truncated and empty fits, multicenter AO/auxiliary image transport, the two K-domain conventions, evaluated QVF densities and the landed FockBuild interface. The complete method plan admits retained fitting plus direct response and composition workspace; plan.fock alone is only the fitting subplan. Thirty-two prechange complete RI states retain arrays, spectra, energies, source/projector IDs and existing resource counts bitwise. Eight isolated benchmark workers measure complete native/public jobs and independently converged references in benchmarks/manta_periodic_mixed_hf.md. Mixed finite-source qualification does not establish fitted KS, physical range fitting, BIPOLE parity or a source-specific symmetry reduction.

Explicit fitted pure/global KS now passes independent libint raw-fitting, Gaussian Fourier/direct and libxc quadrature contractions with NumPy factors and SciPy SCF. Pure RI-J LDA/PBE/BLYP and all three global exact/RI choices cover Gamma/shifted meshes, restricted/unequal/empty spins and strict/BvK corrections. Capped recurrences preserve the evaluated density and frozen projectors. Fourth-order complex congruence energy differences check the functional Fock derivative; the HF half trace is explicitly rejected as an energy reference. Empty PBE/PBE0/B3LYP endpoints keep separate measured minority-potential sensitivity witnesses, with strict occupied-spin and energy checks. This is the existing libxc endpoint limit, not a widened uniform matrix tolerance or a regularized functional.

Truncated/zero fitting spaces, independent multicenter AO/auxiliary image shifts, redundant AO shells and rectangular S/X frames preserve all physical components. Exports cross the actual shared FockBuild and BlockSpaceAction interfaces with method-owned energy and no inferred symmetry certificate. Source changes at the same HCORE state preserve the evaluated-density identity while changing source/Fock identities. Fifty-six saved direct pure/global/range KS states retain arrays, spectra, energies, source hashes and original resource counts bitwise; density/snapshot IDs intentionally adopt the separated version 2 contract.

All five native and selected scientific caps admit exactly and refuse one short before poisoned array copies. Functional/source restrictions remain fail-closed. Twelve isolated benchmark workers, each with five repeated native and selected scientific solves, report complete fixed-source errors and conservative resource models in benchmarks/manta_periodic_ri_ks.md. The selected timing includes provenance construction but no full public job or artifact writing.

Generated fitted KS adds independent integer membership/conjugacy checks for Gamma and shifted meshes, raw-fitting/functional SCF on generated supports, all exact/one-short domain/grid/output caps, capped/truncated states, empty spins, rectangular overlap frames and independent AO/auxiliary images. Owned exports preserve the shared Fock and BlockSpaceAction contracts. Twelve saved generated direct KS states preserve numerical arrays, plans and identities exactly after sharing the factories. The generated/explicit/selected states agree bitwise. benchmarks/manta_periodic_ri_ks_domains.md measures complete source and quadrature construction in twelve isolated workers with five repeats per interface.

Public fitted KS adds independent pure/global functional SCF, explicit J/K selection, complete output-budget boundaries, capped/truncated states and lossless evaluated-density QVF for restricted, unequal and empty spins. Independent AO/auxiliary image shifts, rectangular overlap frames, immutable result/evidence exports, citations, dry runs and artifact-delivery failures preserve the method contract. Sixteen saved public direct KS/RI HF states retain numerical arrays, native plans and identities exactly. The twelve-worker benchmarks/manta_periodic_ri_ks_job.md measures complete native and public jobs, including standard output and QVF, with five repeats each. Physical range fits, BIPOLE/CRYSTAL parity, source-qualified reduction and the remaining approved matrix stay required.

The initial public surface covers the P0 reference boundary and the P1 molecular MD one-electron matrices, streamed 2c/3c fitting tiles and four-center tiles/direct J/K, including erf/erfc kernels and batched AO-pair/auxiliary Fourier panels, multi-density direct digestion and a streamed dense-metric RI-J/K reference, plus bounded sharing across general and partial contraction channels, analytic one-electron basis/charge/origin derivatives and analytic 2c/3c/4c shell-center derivative tiles and a bounded all-electron molecular HF Lagrangian gradient consumer, plus explicit RHF/UHF/RKS/UKS method selection with executed-provider and terminal-density evidence, independent exact/RI range exchange sources and the native XC/VV10 composition. A bounded absolute three-dimensional shell-pair extent/tail envelope and bounded Cartesian multipole translations/interactions and a deterministic streamed bubble partition support a low-level molecular BUPO-J consumer. Independent point-charge Legendre expansions check the multipole algebra, including the compact-support remainder; Gaussian tails remain separate. Small exhaustive inventories prove that near/far blocks partition each unordered shell-pair interaction once, including when candidates are refined. Shared primitive shell-pair moment panels through degree twelve and optional full absolute radial envelopes support streaming forward/reverse consumers; independent polynomial quadrature validates the complete panels. Single-auxiliary panels through l=7 and degree twelve, and shell-centered full/tail extent envelopes, are validated against direct polynomial quadrature and radial gamma integrals. These retain the single-function distribution required by RI-BUPO rather than substituting an orbital product. A Gaussian error model combines compact truncation, exact tail penetration and the full-versus-compact moment correction. Independent libint and analytic Gaussian interactions validate its per-integral approximation ledger, separate from floating-point contraction errors. The paired J consumer passes arbitrary density, linearity, reciprocity, energy-derivative, near-only and strict far-field comparisons against independent dense integral contractions. Its selected RHF/UHF/RKS/UKS adapter confirms the returned density against an independent exact J while preserving the chosen exchange and XC treatment. Loose negative controls must fail that confirmation even when the approximate SCF itself is stationary. Performance benchmarks remain a separate gate; BUPO does not supply exchange or derivatives. Pair-bubble/single-auxiliary interactions reuse the same explicit Gaussian ledger, with independent three-center libint and unequal-Gaussian analytic checks. This establishes coupling-tile admission, not a complete RI-J error or fitting-conditioning certificate. A native paired three-center coupling now streams the same density-independent partition in forward and adjoint passes, with atomic auxiliary leaves and pair bubble translations. Independent libint/libcint comparisons through orbital l=6/auxiliary l=7, full-matrix adjoint tests, exact near limits, signed linearity, positive-metric energy derivatives and exact aggregate resource boundaries validate that primitive. A bounded RI-BUPO J consumer now combines both passes with the existing exact dense metric and fixed positive retained subspace. It reports explicit exact-RI confirmation, separate pass defects, propagated J/energy coupling bounds and discarded-space metric residuals. Independent dense references, duplicate/nearly-dependent and zero-rank spaces, energy finite differences and combined resource boundaries validate this composition. Selected RHF/UHF/RKS/UKS can now use RI-BUPO J with independent exact/RI K and native XC/range exchange. Terminal confirmation rebuilds exact RI-J with the identical auxiliary/metric source; loose stationary and NOITER controls must fail. Receipts include executed paired-source and conditioning evidence. Large streamed metric solves and selected RI-BUPO derivatives remain gates. QVF file and memory producers preserve executed molecular engine receipts in an optional vendor section, including failed exact terminal confirmation, with hash validation, coexistence and source-ownership tests. Selected molecular static populations, bond indices, natural orbitals and dipoles use the executed one-electron provider with content checks and separate work receipts. Selected single-point jobs retain J/K and range/XC sources through dispatch and RKS READ retries, output/citations and failed-confirmation checkpoints. IAO populations and closed-shell IBO/Boys/Pipek-Mezey localization use the selected main, minimal-reference and cross-basis overlaps and dipole moments. Independent libint comparisons, occupied-density preservation, source refusals and QVF exports qualify these property paths. Atomic SAD promolecules, response and other job compositions remain gated. Independent shell-tile sessions admit only provider/workspace resources and one quartet, leaving aggregate work and density buffers to the scheduler; the original ordered J/K plans retain their whole-traversal admission. These interfaces do not complete the remaining job-composition, large-metric, periodic or derivative gates; the manual records what is actually selectable. Symmetry foundations can advance independently, but P5 execution depends on the actual P2/P3/P4 source to be reduced. Derivative primitives can be developed earlier; a user-facing force/stress capability requires the complete P6 gate.

Use incremental, additive changes. Add a manifest classification through scripts/test_gate/update_suite_manifest.py for each new test file. Add user-visible changes as changelog.d/ fragments when behavior ships, not as claims in released changelog sections. A wrong-answer fix or release blocker needs independent review under repository policy; this design does not close any such issue. Before implementing a large refactor, record its narrowed design decision on the tracker. This design pass sends no external messages.

The dense RI metric response now shares the RI positive-eigenprojector factorization. Independent displaced eigensolves validate inverse and projector derivatives, including nonzero discarded modes, degenerate blocks, zero rank and absolute/relative cutoffs. Frobenius adjoints, orthogonal covariance, fixed-density fitting energies, exact factor identity, numerical rank-crossing refusals and resource boundaries are qualified separately from complete nuclear gradients. The cutoff guard is a numerical admission policy, not an interval spectral certificate.

A fixed-density RI Coulomb gradient consumer now streams the fitting RHS and three-center derivative contraction and contracts the exact metric adjoint. Independent libint energy differences validate full and truncated fitting spaces; range kernels, separate orbital/auxiliary ownership, component-wise finite differences, translation/rotation sum rules and aggregate resource boundaries cover the derivative composition. This qualifies the RI Coulomb energy component, not complete SCF forces, exchange derivatives or stress.

The paired-spin RI exchange energy gradient now forms its metric adjoint in bounded raw auxiliary blocks, including discarded modes, and streams the three-center derivative contraction. Independent libint energy and component finite differences, full/erf/erfc kernels, spin additivity, block invariance, zero rank, geometric sum rules and aggregate work/memory admission qualify this component. Receipts match the RI exchange source and spin-density identities and expose the reference schedule’s repeated integral passes. Complete selected RI SCF force composition remains a separate gate.

A complete low-level molecular HF RI Lagrangian now combines independently selected exact/RI J and K derivatives with nuclear, core and overlap Pulay terms. Component-wise independent libint finite differences cover pure and Cartesian, full and truncated metrics; relaxed RHF/UHF differences and independent full-RI force comparisons include auxiliary response. Executed counts prove that an all-RI derivative requests no four-center tiles. The combined admission covers every component before copying input matrices. The generic selected RHF/UHF adapters now retain an immutable auxiliary basis, RI policy and budget. They verify geometry, terminal density and rebuilt RI source identities. Relaxed-energy differences exercise both public and native gradient routes for independent exact/RI J and K selections. Custom same-name auxiliary bases, truncated metrics, caller option mutations, derivative-budget refusals and success-only output receipts qualify context preservation. Selected KS and BUPO derivatives remain separate gates.

The molecular HF geometry source now preserves orbital/auxiliary shell content and atom maps while rebuilding centers at each geometry. Displaced RHF/UHF energies validate all exact/RI J/K compositions; native optimizer walks and complete jobs verify stationary final forces, source identities, final custom auxiliary placement and derivative citations. The source snapshots mutable SCF/fitting options and budgets and uses explicit HCORE starts. Projected READ warm starts and other derivative compositions remain separate gates.

The finite Bloch Fourier primitive now streams translated shell panels across an explicit image list and shares them across ket momenta. Independent Gaussian quadrature, physical libint overlap at zero argument, reciprocal relabeling, global-origin phases, momentum-transfer adjoints and exact physical-support boundaries qualify its convention. The older raw Fourier table’s normalized solid harmonics require an explicit angular conversion before reference comparison, matching the existing RSGDF AO scaling helper. That difference is not hidden by a fitted numerical rescaling; complete periodic/CCM sources still require normalization-consistent method qualification. Aggregate work/storage caps and actual generation/folding receipts cover bounded batches and the opt-in periodic Fourier helper adapters.

The finite three-dimensional reciprocal Hartree consumer now composes this provider over a complete density epoch with explicit nonnegative normalized k weights and unique inversion-closed image/G supports. Tests compare pure and Cartesian results with independent Gaussian quadrature, differentiate the Hartree energy with respect to one complex Hermitian density block, and check Gamma reduction, nonuniform mesh permutation, time reversal, reciprocal relabeling and origin covariance. A normalized scalar Gaussian supplies a closed-form reciprocal-cutoff convergence reference. Existing Ewald J cell/k interfaces agree at the same finite source and physical normalization. Admission covers cumulative kernel work, all live output/input owners and the Python grid generator; low caps fail before payload/grid allocation. Raw imaginary/Hermiticity residuals are checked before legacy real-block conversion. This does not qualify nuclear/K composition, periodic SCF, CCM, periodic BUPO, forces or stress.

The one-electron point-charge provider admits erf/erfc kernels and analytic bra, ket and charge-position derivatives at fixed range. Independent libint erfc nuclear matrices, optional libcint Gaussian-charge matrices through l=6, and high-precision scalar Gaussian integrals check the radial convention. The latter retain relative accuracy checks on erfc tails below the range at which full-minus-erf subtraction rounds to zero. Range partition, translation sum rules, center finite differences, resource boundaries and unchanged default Coulomb source identities gate the primitive. This does not itself admit a periodic nuclear sum or a derivative of the splitting parameter.

Finite overlap and kinetic producers now retain real cell blocks while sharing translated MD shell tiles across all Bloch momenta. Independent libint lattice blocks, l=6 polynomial quadrature and a converged analytic Gaussian chain check the physical convention. Tests cover image-relabeling covariance with regenerated pair-complete supports, reciprocal relabeling, asymmetric/repeated image domains, raw residuals, exact cutoff boundaries and aggregate admission. The explicit cell adapter preserves the standard LatticeMatrixSet in dimensions one through three and emits executed source/work receipts. It does not enable complete periodic SCF or symmetry reduction on its own.

The finite three-dimensional nuclear operator now composes erfc point-nucleus images, damped reciprocal attraction and the positive overlap correction for zero-average G=0 treatment. Its physical nuclear sphere uses the existing midpoint-centered selector. Independent libint short-range blocks, Gaussian quadrature reciprocal blocks, analytic overlap corrections and a converged scalar plane-wave series check each component and the total. Tests exercise split convergence, basis/nuclear image relabeling, origin shifts, complex Bloch folding, l=6 screened attraction, actual image/primitive-charge counts and aggregate resource boundaries. The explicit lattice adapter agrees with the existing Ewald operator at matched finite domains and emits execution evidence. This qualifies the operator, not complete periodic SCF, exchange, forces or stress.

A finite reciprocal full-Coulomb exchange consumer now accepts explicit per-transfer Q supports and an immutable all-k/all-density snapshot. Its B-dagger-D-B convention agrees with independent Gaussian quadrature, analytic scalar Gaussian sums and the existing negative-source-phase contraction after physical AO normalization. Reversed transfer supports are checked so that the weighted density response is self-adjoint; perturbing one k density verifies the exchange-energy derivative and changes responses across the mesh. Separate spin/trial channels share Fourier work but retain separate responses/identities. Tests cover image-gauge density pushforward, reciprocal relabeling, arbitrary complex Hermitian densities, raw residuals, batching and aggregate admission. The explicit source omits Q=0 and adds no Madelung correction. It does not qualify singularity-corrected exchange or complete periodic SCF.

The same consumer now supports explicit erf/erfc kernels. A positive-range erfc request includes its finite zero mode only when supplied in the domain; full Coulomb/erf retain the omitted-zero convention. Independent quadrature, energy derivatives, high-precision positive-tail tests and a scalar Gaussian real/reciprocal Poisson identity qualify this distinction. Kernel/range and actual zero-mode work are recorded; no singularity correction is inferred.

The native spin Fock consumer now freezes H+J[D_alpha+D_beta]-a*K[D_spin] and the corresponding electronic-energy components in one owned epoch. Independent quadrature, spin/k energy derivatives, restricted factors, time reversal and atom-image gauge tests qualify the composition. Source, density and evaluated-payload identities remain distinct; aggregate admission reserves both component peaks and enclosing owners before copying inputs. The H matrix is explicitly supplied and authenticated, not asserted to be a MANTA nuclear source. This is a fixed matrix snapshot; complete physical periodic SCF and correction policies remain separate qualification stages.

The neutral all-electron three-dimensional RHF driver now composes physical S/T/V, point-charge Ewald energy and the whole-mesh reciprocal J/K source. The bounded numerical solver retains the last evaluated density/Fock pair, checks the occupied/virtual gap and projected eigenproblem, and requires a fresh terminal Fock/energy confirmation. Explicit strict zero-mode omission and a separately admitted BvK probe-charge Ewald correction are qualified against independent quadrature/libint components and optional refined PySCF AFT calculations. Full regular Gamma, multi-k and shifted meshes are covered; cutoff convergence remains distinct from SCF convergence.

The cutoff factory admits physical AO-pair image spheres and certified reciprocal kinetic-energy spheres before allocation, checks reversed momentum-transfer conjugacy, and preserves independent domain/source identities. Its selected run_periodic_job adapter now provides separate engine/J/K controls, explicit domains, output memory reservations, standard artifacts and executed provenance. QVF preserves evaluated densities on converged and capped exits and transports independently displaced AO shells into the archive’s nuclear-center gauge. Capability refusals and output delivery failures retain the shared job lifecycle semantics. Native-source and complete-job benchmarks separately report numerical error, time, work and memory. General periodic KS, RI/BUPO, reduced-dimensional boundaries, CCM, qualified symmetry reduction and periodic derivatives remain open stages; these RHF results do not qualify them.

The physical RHF driver also retains scalar snapshots of every evaluated iteration through the bounded solver’s existing observer contract. Native copies and Python/QVF records are admitted against the maximum iteration count; matrix histories are not retained. Independently capped runs check each energy/residual/gap snapshot, and confirmation calls do not create extra iterations. Selected jobs deliver the complete history after the solve while retaining the evaluated terminal density separately.

The bounded UHF numerical leaf is now a separate fixed-spin solver with D[2,K,n,n], shared physical S/H[K,n,n], independent alpha/beta occupations and a single whole-density callback per iteration. It shares the RHF leaf’s audited Jacobi, overlap and metric routines. Complex NumPy fixed-point references qualify unequal/empty spin populations, spin exchange, AO phases, per-spin band gaps, metric idempotency and last-evaluated-density exits. Exact-cap and one-byte-short cases check owned and enclosing resource budgets. The explicit-support physical UHF driver now composes this leaf with the neutral 3D all-electron S/T/V and nuclear Ewald source, whole-mesh spin J/K, and separately admitted BvK correction. Independent quadrature/libint UHF fixed points validate empty and nonempty beta channels, Gamma and shifted meshes, component energies, shell-image gauge transport and same-density confirmation. Physical RHF/UHF share the spin-independent source receipts; the full UHF identity additionally binds its fixed spin populations. Refined Gamma and shifted-mesh hydrogen cases also compare S/H, both spin densities and Focks, nuclear and total energies, and the Madelung constant against optional PySCF AFT KUHF. The reference uses explicit fixed-spin occupations to avoid its empty-beta diagnostic bug, with supercell electron counts set explicitly. These are finite-basis checks, not spin-stability or basis-convergence claims. Cutoff-generated UHF now shares the certified geometric source with RHF and admits a separate complete spin-aware physical plan. Independent integer enumeration, spin SCF, shell-image transport and exact-budget refusals qualify this route. Selected public UHF jobs now deliver separate evaluated spin densities, owned spin orbitals/occupations, complete scalar spin history and standard artifacts. Independent job-level integral/SCF comparisons and QVF readback cover empty beta, unequal populations, capped states and shifted AO gauges. Job resource admission and delivery failures preserve the shared lifecycle. Pure-KS jobs are qualified below; general KS and later stages remain open.

A separate joint functional numerical contract now reuses the bounded restricted/unrestricted iteration kernels while preserving the HF APIs. The provider supplies potential and interaction energy from one evaluated D. Nonlinear polynomial fixed points and energy derivatives distinguish this from the HF half-trace formula; capped, cancelled and missing-output cases check epoch consistency. This is groundwork for periodic KS, not physical XC quadrature or selected KS job qualification.

2. Independent reference hierarchy

Agreement with one backend does not establish correctness if both share a normalization or contraction bug. Use all applicable levels:

1. Analytic ssss, overlap, kinetic, point-charge and Gaussian moment formulas; high-precision one-dimensional Boys quadrature and small Gaussian Coulomb cases independent of the production recurrence.

2. Existing libint integral tiles, with a separately checked AO convention; small independent NumPy contractions over full tensors, not the same MANTA digestion loop reused as an oracle.

3. Existing unreduced native/Python lattice and reciprocal builders under matched supports, weights, metric ranks, background and q=0 conventions. Distinguish finite-source parity from convergence to an infinite source.

4. Independent external programs only in optional tests or regression subprocesses, never in python/vibeqc/ or cpp/. Pin basis coefficients, pure/Cartesian order, pseudopotentials, auxiliary basis, exchange treatment, numerical cutoffs and gauge. A converged SCF comparison without those matches is a scientific comparison, not an implementation identity test.

For periodic Coulomb use an independently accumulated high-accuracy Ewald or reciprocal reference; validate nuclear, J and combined electrostatic energy separately. A direct bare lattice cutoff is not a valid 3D Coulomb reference without its surface/neutrality convention. For CCM use explicit small-domain weighted sums as oracles; a generic periodic program cannot certify a different finite Hamiltonian. Published lattice-sum numbers in Kudin 2004 Tables I-II provide an additional convention-converted witness.

Existing tests to inspect and extend, preserving their established thresholds:

Layer	Existing files
Molecular primitives and fit	tests/test_integrals.py, tests/test_df_core.py, tests/test_df_rhf.py, tests/test_df_uhf.py, tests/test_df_gradient.py
Frozen operators	tests/test_operator_contract.py
Periodic finite kernels/resources	tests/test_periodic_gaussian_source_context.py, tests/test_periodic_gaussian_metric.py, tests/test_periodic_gaussian_three_center.py, tests/test_periodic_gaussian_fock.py
Whole-mesh and GDF	tests/test_periodic_fock_multi_k.py, tests/test_rsgdf_shared_q_batch.py, tests/test_gdf_cross_route_parity.py, tests/test_periodic_gdf_memory_preflight.py
BIPOLE identity and final density	tests/test_bipole_jk_domains.py, tests/test_bipole_domain_contract.py, tests/test_bipole_final_density_phase.py, tests/test_pbc_bipole_multik_ewald_split.py
CCM/chi	tests/test_ccm_bloch_reference.py, tests/test_ccm_direct.py, tests/test_periodic_ccm_neutral_bloch_adapter.py, tests/test_chi_transform_inputs.py, tests/test_chi_fitted_controls.py, tests/test_chi_hf_reference.py
Symmetry	tests/test_symmetry_shared.py, tests/test_symmetry_integrals_reduced.py; the symmetry workstream’s retained-GDF tests after integration

3. Numerical acceptance suite

The numbers below are proposed initial small-fixture gates, to be confirmed with independent references. They do not relax existing tests or promise these tolerances for arbitrary ill-conditioned inputs. Fixtures have bounded norms and non-singular retained spaces. If a gate fails, establish conditioning/reference error before proposing a change; do not silently widen it. Publish absolute max, absolute Frobenius and scaled Frobenius residuals.

Test family	Required cases and measurements	Initial gate
Normalization and integral tiles	All s/p classes and representative classes through advertised l; mixed pure/Cartesian, same-center, far/diffuse/tight, signed and deep contractions, ghosts; 2c/3c/4c; every coincidence/permutation	Element error <= 1e-11 + 5e-12 * abs(reference) a.u.; additionally report norms and analytic ssss/S diagonals. Higher-conditioning cases use high-precision witnesses.
Range and moment/FT kernels	omega near zero, intermediate and large; full=erf+erfc; analytic finite SR Q=0; moments at several origins, translation composition and derivative orders; FT conjugation and q wraps	Same integral gate; zero/large limits and exact low-order identities tested independently
Arbitrary-density J/K	Seeded PSD, Hermitian indefinite, sparse, random complex, spin-separated and small delta densities; batches of 1/2/8 targets; response non-Hermitian mode separately	Max error <= 2e-10 Eh for fixed bounded fixtures versus independent full contractions; linearity/adjoint and raw Hermiticity <= 1e-11 in their stated norms
Energy consistency	Central difference of quadratic J/K energy with nonsymmetric-in-space Hermitian perturbations; density-pair reciprocity; spin factors and separate exchange ranges	Directional derivative agrees to 1e-8 Eh on the step-size plateau; cross-density bilinear residual <= 1e-10 before any symmetrization
BUPO-J	Exact near-only, far-only admitted pairs, translated and rotated geometries, signed contractions, diffuse always-near bucket, repartitioned trees, reordered tasks	At strict target, max delta-J <= 1e-8 Eh and absolute delta-EJ <= 1e-8 Eh for small fixtures; tighten each approximation until the same-source analytic limit is recovered
RI-BUPO-J	Exact B/V baseline; approximate forward only, adjoint only and paired passes; rank-degenerate/near-threshold auxiliary spaces; small/large solve residuals	Same BUPO gate relative to exact RI; forward/adjoint dot test <= 1e-11 relative. Separately report exact-RI minus exact-4c error, never charge it to BUPO
Incremental builds	Alternating and cancelling deltas, different alpha/beta slots, restart, threshold/rank/kernel/tree changes, at least two rebuild intervals	Final full build agrees with independent nonincremental map within its allocated error; forced drift prevents convergence and reseeds state
Periodic finite-source parity	Gamma, 2x1x1, 3x1x1, 2x2x2 and shifted meshes; non-TRIM complex densities; one k perturbed at a time; nonuniform normalized quadrature outside finite-character tests	Raw J/K max error <= 1e-9 Eh for small matched sources; component energies <= 1e-8 Eh/cell; no final averaging to conceal phase defects
Periodic convergence	Independent real/G/eta, near-cell, pair-image, multipole order and mesh ladders; diffuse/skew ionic cells, neutral and explicitly compensated charged controls	Same-source errors meet the specified budget; successive independent converged controls change total energy < 1e-7 Eh/cell and J/K < 1e-7 Eh before claiming convergence for that fixture
Cell identity/gauge	Shift atom by lattice vector; change unit-cell origin; equivalent supercell folding and k labeling; G=0 background and nuclear self removal; full versus SR exchange seam	Covariant matrices and invariant matched energies satisfy the periodic gates; expected potential-zero changes are identified component by component
CCM identities	Explicit weighted small sources, WSSC ties on odd/even meshes, all route names, neutral real/Bloch Fourier equivalence, chi finite characters and fitted projectors	Same-source matrix/energy gates; intentionally different sources must remain distinguishable; unsupported weight factorization/dimensions refuse
Symmetry	Proper/improper operations, stabilizers, fractional translations, scalar TR, Cartesian metric, dense auxiliary gauges, broken-state trial densities, orbit-adjoint contraction	Reduced/unreduced max delta-J/K <= 1e-10 Eh on bounded fixtures, separately from approximation error; every deliberately invalid source/rank/action is rejected or runs identity-only
End-to-end SCF	H2O, stretched/open-shell molecule, signed/general/diffuse basis; ionic and covalent distorted solids; each experimental source separately; HF, pure KS, hybrid, screened hybrid, smeared metal where supported	Same basin/occupations and original convergence tolerances; total energy <= 1e-8 Eh (per cell for periodic fixtures), component errors reported, density subspaces compared in S rather than arbitrary orbital gauges
Gradients	Distorted molecules/cells, selected exact/RI/BUPO energy, multiple displacement steps; total and component finite differences; tree/rank/support crossings	Max force error <= 3e-6 Eh/bohr on a stable FD plateau; translation/rotation sum rules; separate reports for crossing cases and nonvariational sources
Stress	Hydrostatic and shear strains, skew cells, fixed fractional k, nuclear/background/seam terms, finite-temperature free energy where applicable	Max error <= 3e-6 Eh/bohr^3 for 3D fixtures; dimensional units for 1D/2D; compare several strain steps and cutoff ladders

Convergence ladders must vary one approximation at a time and then test their combination. Finite-difference energy noise must be well below the targeted force times displacement (or stress times volume and strain). Smearing tests compare the appropriate free energy and occupations. Degenerate orbital rotations are assessed as subspaces; energy ordering alone is not a state identity check.

For every budgeted path test exact-cap acceptance, one-byte/work-unit-short refusal, integer overflow, cancellation, invalid dtype/layout, NaNs and stale source identity before expensive allocation. Verify that tile views cannot outlive their workspace and that mutating a builder after snapshot creation does not change a captured operator. Concurrent density mutation is prohibited; use ownership or the existing checked borrowed-input pattern. Force a fallback and verify its source identity, memory admission and output counts.

Initial local test ceilings: analytic/tile fixtures 60 seconds and 512 MiB per case; small J/K and periodic cases 15 minutes and 4 GiB per process. These are proposed admission limits, not measured runtimes. Medium numerical and scaling cases need a separately sized plan after small gates pass. No full native lane, large all-q inventory or large supercell run is required by this documentation change. Actual implementation tests must rebuild and verify the loaded native extension belongs to the edited checkout.

4. Benchmarks with attributable speedups

Benchmark the requested scientific accuracy, not only a kernel’s throughput. Use reproducible public input definitions, pinned source/build/compiler/BLAS, CPU/GPU model, thread/rank settings and numerical controls. Record operational logs, machine identifiers and absolute paths outside the product checkout. Only sanitized portable summaries and approved fixtures belong in the tree. Acquire licensed literature SI inputs on demand unless redistribution has been reviewed. SI geometry presence does not confer a license to bundle it.

Workloads should cover:

• Low-l segmented molecules where libint can win; deep general contractions; high-l quartet and 3c classes; multiple response right sides.

• Compact water/organic clusters, extended glycine chains and a diffuse charged molecule. Vary shape and basis separately from atom count.

• Gamma and multi-k covalent/ionic solids, anisotropic cells and slabs/wires only after their source gate. Vary cell size, k mesh and support separately.

• Small and growing finite sources for each CCM/chi route, including cases where weight structure prevents efficient BUPO aggregation.

For each workload run these matched variants, including no-win cases:

Variant	Provider	J algorithm	Symmetry	Question answered
A	libint	existing exact/direct or RI	off	Baseline including existing optimized digestion
B	libint	new tile/consumer exact or RI	off	Interface/digestion overhead or benefit
C	MANTA	same exact or RI algorithm as B	off	Integral-provider contribution
D	MANTA	BUPO or RI-BUPO matched to C	off	Coulomb approximation/traversal contribution at measured accuracy
E	MANTA	same as D	qualified reduction	Symmetry contribution for the same numerical source
F	libint near field	same BUPO plan as D	off	Provider versus BUPO interaction/crossover

Measure A/B, B/C, C/D and D/E separately; show A/E only as total observed benefit. Also compare to the best existing admissible route. Approximate sources with different RI/K choices do not form a speedup ratio without showing their accuracy difference. End-to-end timings are not products of microbenchmark speedups.

Each record contains setup and total wall/CPU time, SCF iterations and final confirmation time; basis/extent/tree setup; E/R/Boys formation; primitive and contracted transforms; screening; near-field integrals; moments/M2M/M2L/L2L; metric setup/solve; RI forward/adjoint passes; J/K digestion; reciprocal work; symmetry planning/transport; XC; eigensolve; output; synchronization, I/O and host/device transfers. Report peak process-tree RSS and device memory, measured workspace high-water mark, per-rank/worker replication and spill volume.

Actual work counts include candidate/retained shell and primitive pairs, quartets/triples/metric tiles, screened terms, executed GEMMs and recurrence classes, bubble visits/accepted interactions/orders, near-cell and reciprocal vectors, q/k contractions, fit iterations, symmetry representatives and reconstructions, incremental/full rebuilds and fallback tiles. Do not present planner cardinalities as executed work. Explain time components that overlap instead of adding overlapping timers.

Run one warm-up and at least three measured repeats for stable medium cases; report median and range, with cold setup separately. Longer cases may use fewer repeats with that limitation visible. Confirm accuracy on the timed run’s actual output, including J/K, energy components and terminal density. Control thread counts and node contention; an isolated rerun distinguishes a timeout from a numerical regression. Show absolute seconds and memory even when a speedup is below one. Fit scaling only over the measured regime with fixed accuracy, and disclose dense metric/solver bottlenecks and diffuse near-field growth. No ORCA speedup is transferred to vibe-qc.

5. Decisions and review checkpoints

Adopt now as design constraints: explicit provider selection, separate J/K algorithms, fixed solver snapshots, source-specific periodic/CCM identities, bounded production storage, paired RI approximation, unreduced reference and visible provenance. These do not change a public API or an existing source.

Before the relevant implementation stage, review the proposed public option names/schema, qualified l/derivative envelope, calibrated accuracy profiles, periodic tail representation, compatible CCM weight factorizations and the large-metric solver/preconditioner. A new hard dependency, licensing decision, public API break or physical-source/default change belongs to the maintainer. Ordinary additive provider implementation and focused validation can proceed within the agreed stage. Preserve the libint reference throughout the rollout.

Periodic direct-Bloch local-exchange increment

The finite local-exchange source is qualified separately from physical KS SCF. tests/test_manta_periodic_local_exchange.py compares complex direct Bloch densities and potentials with independently evaluated Gaussian polynomials and libxc LDA exchange. Its gates cover weighted k/spin energy derivatives, restricted spin scaling, empty spin, shell-image gauge changes, source identity/ownership, image and grid refinement, and exact/one-short resource admission. No real inverse-Bloch density assumption is made.

This leaf admits explicit 3D image/grid supports and l<=4; it alone does not close the periodic KS capability row. The following increments qualify pure LDA/GGA physical composition, grid/domain generation and public job outputs. General XC, production accuracy profiles and other unqualified stages remain required.

The subsequent native LDA/GGA component shares that finite Bloch support and adds bounded AO gradients, spin-density invariants and their exact adjoint. Independent Gaussian polynomial derivatives and libxc point evaluations, complex real/imaginary energy differences, mixed-spin gradients, restricted limits, empty spins, image gauges and exact resource boundaries qualify it. The finite physical periodic KS composition below consumes this joint functional energy contract, preserves the same k weights and confirms the terminal evaluated density against the complete finite physical source.

The explicit physical pure-LDA/GGA KS driver now composes the qualified Hartree and semilocal sources with physical S/T/V and nuclear energy. The same-density joint energy contract applies to both restricted and fixed-spin unrestricted occupations. Independent Gaussian/libxc contractions and SciPy eigensolves check full SCF, capped evaluated epochs, separate energies and final physical confirmation. Exact/one-short aggregate budgets include both sources and the functional solver’s retained energy scalar. The finite Hartree pair support and canonical XC image support remain distinct; grid charge, XC energy and XC potential require independent refinement. Public pure-KS jobs have their own acceptance below. General functional profiles are not inferred from this finite-source increment.

The explicit global-hybrid adapter separately adds full-Coulomb exchange to the semilocal component on the same evaluated density. Resolved method components supply the coefficient; the pure API still refuses hybrids. Independent Gaussian pair Fourier integrals and AO gradient contractions, libint one-electron matrices, libxc integrands and SciPy fixed-point solves check PBE0, B3LYP and custom LDA/GGA hybrids. Both zero-mode policies have independent probe-charge and spin-factor witnesses. Complex congruence energy differences check the Fock derivative, including k weights and conjugation. Capped epochs, source/image gauges, rank truncation and all five exact/one-short aggregate resource limits pass. Separate empty-spin PBE0 and B3LYP tests measure libxc endpoint sensitivity without changing runtime regularization or claiming uniform minority-potential accuracy. Full-matrix empty-spin acceptance uses a well-conditioned B88/LYP global hybrid.

Exported hybrid snapshots were checked against the actual shared FockBuild draft, including physical S/X frames, owned evaluated densities, separate method energy components, solver-modified operator parentage and next orbital projectors. This qualifies export compatibility with that inspected draft, not a landed runtime adapter or an in-loop Python allocation contract. No symmetry-reduced or BIPOLE producer is admitted by this evidence. General periodic accuracy qualification remains required. Explicit range separation has the separate acceptance below. Reproducible finite-source timing and independent errors are recorded in benchmarks/manta_periodic_ks.md in the repository.

The explicit range-separated adapter now independently qualifies complete finite-source RKS/UKS for HSE06, CAM-B3LYP, LC-omegaPBE, omegaB97X, mixed ranges and empty-spin LC-BLYP. The reference uses fixed published coefficients, independent Gaussian Fourier panels and analytic screened kernels, libxc AO contractions and SciPy SCF. Tests isolate the finite erfc zero and its k weight, the long-range coefficient of the Madelung term, and the absence of that term for screened-only exchange. Both requested zero-mode choices, complex density energy derivatives, capped evaluated states, next projectors, closed-shell limits, retained-rank solves, cancelled components and source-image gauges pass. All five exact and one-short aggregate budgets include every kernel and the expanded transfer source; invalid numerical arrays remain unread after a resource refusal. Unsupported functional families stay gated.

The actual shared FockBuild/OperatorFrame draft also consumed nine exported range-separated states, including early capped exits and converged results. All 45 spin/k operators preserved physical frames, evaluated density, component energy, immutable ownership and solver-modification parentage. Independent generalized spectra and next projectors agree. This qualifies export compatibility with that draft, not a landed in-loop adapter or a symmetry certificate. Canonical Coulomb-plus-erfc zero-mode semantics are explicit; no equivalence to another finite-mesh singularity prescription is inferred. Timings and errors are recorded in benchmarks/manta_periodic_range_separated_ks.md. Range-separated cutoff/grid generation and public output have separate acceptance below; source refinement and production accuracy profiles remain required. BIPOLE/CRYSTAL and symmetry gates retain their own sources.

The range-separated cutoff and joint-grid factories reuse the qualified HF/global-hybrid geometric producer, with complete range-separated admission. Independent integer spheres check Gamma and shifted skew-cell G/Q support, exact boundaries and distinct pair/XC image domains. HSE06, CAM-B3LYP, mixed ranges and empty-spin LC-BLYP match explicit composition bitwise and an independent Gaussian/libxc and SciPy SCF reference. Every erfc leaf retains its finite diagonal zero; domain receipts describe the common nonzero base. Both spin conventions, singularity choices and grid partitions, capped evaluated states, AO-image gauges and standalone receipt lifetimes pass. All five jointly exact and one-short caps precede payload copies or invalid grid generation. The grid, geometric source and all exchange components remain charged together. A further nine-state/45-block check against the current FockBuild draft preserves domain/quadrature provenance and evaluated state ownership. Public jobs have separate acceptance; no symmetry admission follows from these factories.

Hybrid cutoff and joint-grid factories now have independent acceptance too. They reuse the HF geometric planner and certified reciprocal generator with complete hybrid admission; no HF solver proxy is used. Independent integer enumeration covers skew Gamma and shifted multi-k G/Q spheres, exact cutoff boundaries, pair-complete source translations and distinct canonical XC images. Identical pair/G/Q supports retain the same child geometry identity as HF. RKS, UKS, unequal/empty spins and both zero-mode policies match explicit source composition and independent full SCF, Fock and individual energy terms. Both atomic partition choices and capped evaluated states are covered. All five whole-request exact/one-short caps precede enumeration or poisoned array copies, and a coincident-center fixture proves joint refusal before partition generation. Grid/source ownership is charged through the full hybrid solve. These tests do not establish automatic grid accuracy, BIPOLE parity or a symmetry-reduced producer. Public output and QVF qualification follows below.

Pure-KS cutoff generation now reuses the pair-complete direct source and certified reciprocal source, with a separately selected canonical XC image sphere. Count-only admission bounds all traversal, source ownership and the complete physical KS solver; no exchange-transfer tables are constructed. Independent integer spheres, generated/explicit equality, independent SCF, AO-image gauge transport and exact/one-short budgets qualify these domains. The standalone cutoff interface keeps caller-supplied quadrature; the joint producer below supplies it explicitly, including in selected pure-KS jobs. The empty-beta PBE witness also requires a separate minority-potential conditioning check: the observed libxc fp64 endpoint response amplifies one-epsilon input perturbations even when energy and occupied-spin results agree tightly.

A bounded periodic atomic quadrature adapter now qualifies a finite 3D source: shared generic radial/Lebedev rules and periodic Becke/Stratmann partitions, canonical nuclei and folded points, owned exports and count-only work/memory admission including OpenMP neighbor scratch. Independent radial, angular and integer-image partition reconstruction checks the generated weights. Separate Gaussian/libxc contractions check the same-grid energy/potential, and two independent tensor-cell resolutions qualify the refinement reference. Observed grid errors remain explicit: a larger grid is not an automatic accuracy qualification, and partition-radius refinement matters separately. This does not close automatic profiles or low-dimensional grid rows.

Joint grid/domain/KS admission now reserves every retained grid owner and explicit source copy, subtracts grid work before downstream admission, and checks the whole request before generating numerical points. Independent finite-source SCF and bitwise equality with explicit composition cover RKS, UKS, empty-spin BLYP, both partitions, converged and capped epochs. Exact simultaneous and one-short memory/shell/primitive/scalar/kernel limits cover the joint plan; invalid partition generation is not reached after aggregate refusal. Source receipts and histories retain their existing semantics.

Selected public RKS/UKS jobs now use the joint grid/domain/SCF source with explicit finite XC controls. Job-level independent Gaussian/libxc contraction and SciPy SCF references cover PBE restricted and fixed-spin states, the restricted UKS limit and empty-beta BLYP. QVF validation and exact density readback cover the actual evaluated state, including capped epochs; separate generalized eigensolves check terminal orbitals without reconstructing energy from them. Gauge tests distinguish source-frame arrays from nuclear-frame QVF and require a changed frame identity. Owned exports remain independent.

Before numerical generation, selected KS jobs reject unsupported functional components, exchange sources, BIPOLE controls, symmetry reduction and each insufficient resource cap. Grid-only dry runs and artifact delivery failures preserve the standard lifecycle. Functional/grid citations, child source identities and raw component energies survive output. Whole-job benchmarks include artifact writing and report independent errors with modeled memory. This is finite-source job acceptance, not BIPOLE/CRYSTAL parity, a certified symmetry source, or general functional/grid accuracy qualification.

Selected global-hybrid RKS/UKS jobs now pass independent Gaussian/libxc contractions and SciPy SCF checks under both exchange zero-mode conventions. PBE0, B3LYP and custom LDA/GGA components cover Gamma, shifted meshes, restricted, unequal and empty spins. Method-scaled exchange, its separate correction and the unscaled K child receipt survive manifest, structured log and QVF delivery. Component sums refer to the evaluated density, including capped exits; independent generalized eigensolves verify the separate next projector. Nuclear-frame QVF readback and source-frame Fock/gauge checks preserve that distinction. All simultaneous resource caps pass exactly and refuse one short after reserving job and provenance owners. Unsupported functional components, BIPOLE and symmetry requests fail before generation; dry-run and artifact-failure checks preserve the existing output lifecycle. The public-job benchmark times complete artifact delivery and independently checks SCF, every energy component and lossless QVF density. Finite-source acceptance does not establish BIPOLE/CRYSTAL parity or a symmetry certificate.

Selected public range-separated RKS/UKS jobs now pass independent full SCF, same-density energy/Fock, evaluated-epoch QVF and component-provenance checks. HSE06, CAM-B3LYP, LC-omegaPBE, omegaB97X, distinct-range composites and empty-spin LC-BLYP cover Gamma and shifted meshes. Cancelled descriptors retain complete method admission. Screened-only jobs report the requested convention separately from its inactive actual correction, with identical source/D/F under changes to inactive controls. Signed component energies, raw child K receipts and complete semilocal requirements survive JSON output. All five exact/one-short job budgets include component receipt and serialized owners; unsupported kernels and sources fail before generation. Capped states, image gauges, immutable public exports, dry runs and delivery failures retain their contracts. Nine public states also pass the actual FockBuild draft adapter check with full delivered provenance. Method-owned convergence norms and confirmation criteria remain native; a generic operator diagnostic does not replace them. Whole-job benchmarks include standard artifact delivery.

The frozen-density grid assessment now has its own native aggregate admission and an explicitly partial hybrid entry point. Independent Gaussian/libxc contractions check global and range-separated LDA/GGA components, including multiple ranges and empty spin. Complete hybrid builds at the same HCORE density independently confirm that only the semilocal Fock/energy changes when the grid changes. Pure-entry refusal, complete hybrid admission, cancelled descriptors, source/gauge ownership and exact/one-short aggregate resource caps retain their gates. Exchange and singular corrections remain with the complete method, and partial potentials are never labelled Fock.

The pure grid assessment also has independent contraction checks. LDA/PBE restricted and unrestricted states, BLYP, empty-beta BLYP and complex non-TRIM k matrices check signed XC energy, each spin population, the reference potential, and every spin/k maximum-entry and Frobenius difference. All five resource caps pass exactly and refuse one short before copying invalid numerical inputs. Tests cover count overflow, retained reference and external storage, batch-independent source identities, separate fixed-density identity, owned exports and AO-image gauge transport.

Its benchmark separates same-grid correctness from reference convergence. The chosen atomic reference still has measurable energy and potential errors against independent tensor-cell integration. The optional physical HeH/PBE probe holds a confirmed finite-source UKS density fixed across four atomic grids and three independent tensor resolutions. It tests reference adequacy, not another SCF sequence; a stable charge integral cannot qualify a changing minority-spin potential. No automatic accuracy profile follows from these diagnostics. Image, functional, boundary, BIPOLE parity and source-specific symmetry gates remain independent.

The harder HeH/PBE reference is now also tested with independent adaptive cell integration. The benchmark uses Genz-Malik and two product Gauss-Kronrod rules with explicit subdivision caps; energy, populations and all real/imaginary spin/k potential entries participate in adaptation. The vector witness is checked against same-point native contractions and complex density energy derivatives, including restricted LDA/PBE and empty-beta BLYP. Actual convergence status and inter-rule differences remain distinct from estimated errors. This strengthens the fixed-fixture reference without claiming an automatic grid profile or changing a product quadrature source.

Physical range fitting sources

tests/test_manta_periodic_range_fitting.py compares dense and streamed erfc/erf metrics and couplings with libint short-range and independent Gaussian Fourier references, including signed angular contractions. It checks physical/numerical range separation, finite-zero ownership, the complementary-kernel identity, image gauge and transfer conjugacy, fixed RI contractions, source/frame/density identity separation, immutable controls and all five resource refusals. Saved default Coulomb matrices, identities and counts remain unchanged.

Four complete finite Gamma SCF diagnostics combine the native fixed Coulomb-J, physical-range K and semilocal XC components outside the bounded driver. Their full functional energies and states are compared with independent libint/Gaussian/NumPy responses and SciPy fixed points. These deliberately specified finite functionals have no matched ORCA or CRYSTAL reference; external parity is untested. They do not qualify production fitted range KS composition, public jobs, source reduction, derivatives or a converged accuracy profile. The measured source and calculation results are in the range-fitting benchmark.

tests/test_manta_periodic_ri_range_fock.py extends physical-range acceptance to one, two and four k points, including zero/nonuniform weights, signed exchange coefficients, positive truncated and empty projectors, weighted variation and reciprocity, image transport, exact resource boundaries and screened-J refusal. Four complete shifted two-k functional SCF diagnostics independently compare the whole-mesh native range K with per-q libint/Gaussian/NumPy contractions. They retain separate bare-Coulomb J and actual functional energy. The original Coulomb Fock baselines remain bitwise unchanged. See the whole-mesh range benchmark.

Bounded fitted range KS

tests/test_manta_periodic_ri_range_ks.py independently reconstructs each Coulomb/erfc metric and coupling using libint and Gaussian Fourier references, retains fixed NumPy projectors, and compares complete native SCF states with SciPy fixed points. RI-J/exact-K, exact-J/RI-K and RI-J/RI-K cover Gamma and shifted two-k RKS/UKS, HSE06, CAM-B3LYP, mixed physical ranges, other long-range hybrids and empty beta spin. Tests retain actual functional energy, signed and cancelled terms, finite screened zero and method-owned singular correction.

Further witnesses cover functional derivatives, capped evaluated epochs, truncated/empty metrics, rectangular overlap spaces, immutable exports, all five exact/one-short resource caps before payload copies, and selected Python provenance admission. Shared Fock snapshots and complex multicentre AO/auxiliary image transport keep the operator and symmetry seams qualified for these sources. Sixteen saved prior fitted pure/global and direct range KS states, plans and identities are unchanged. The bounded range KS benchmark records 18 isolated complete calculations. No matched ORCA/CRYSTAL finite-source reference is available; external parity and domain/grid convergence remain untested. Generated range-fitting methods are qualified separately below; public jobs follow below, while source reduction still needs its own acceptance.

Generated fitted range KS

tests/test_manta_periodic_ri_range_ks_domains.py checks integer reciprocal spheres in skew cells, full transfer reversal and kernel-independent geometry. Complete generated RKS/UKS on Gamma and shifted meshes covers HSE06, CAM-B3LYP, mixed physical ranges, exact/RI J/K choices and empty beta. Generated and explicit matrices, spectra, energies and physical-source identities agree bitwise. Independent libint/Gaussian/NumPy/libxc/SciPy SCFs compare complete evaluated matrices, functional components and grid populations.

All five joint resource boundaries, selected output reservation, invalid-source refusal before payload generation, owned arrays, geometry/projector identity separation, shared Fock snapshots, rectangular overlap frames and complex multicentre AO/auxiliary image transport are exercised. These tests do not grant source-qualified symmetry skipping. The generated range KS benchmark records isolated complete native and selected calculations. Grid/cutoff convergence and external ORCA/CRYSTAL parity remain unqualified.

Public fitted range-KS jobs

tests/test_manta_periodic_ri_range_ks_job.py compares complete public RKS/UKS calculations with native generated sources and independent functional SCFs. HSE06, CAM-B3LYP and mixed ranges cover Gamma/shifted meshes and strict/BvK conventions. Separate witnesses cover LC-wPBE, empty beta, truncated/empty projectors, capped evaluated states and every exact/one-short resource limit. Unqualified selections refuse before generation or calculated artifacts.

Returned results, manifests, logs and QVF retain the same energy, density, frame and per-kernel source/projector evidence. QVF schema validation and bitwise density recovery, complex AO/auxiliary image transport, rectangular overlap frames and immutable shared Fock states guard the common interfaces. Dry runs and synthetic delivery failures check writer lifecycle. The public range-KS benchmark measures complete jobs and records independent comparisons, output sizes and resource reservations. These finite-source calculations do not qualify an accuracy profile, external parity, source reduction or derivatives.