Full MANTA implementation work tracker

Development tracker #617, MANTA full implementation alongside libint: capability and claimable work tracker, covers the full approved design. MANTA stays an optional provider alongside libint. Provider selection, Coulomb algorithm and exchange approximation remain independent; BUPO supplies J, not K. The full goal remains unfinished.

The engine capability matrix, scientific contracts and implementation and acceptance plan retain their technical requirements. Issue numbers below refer to the development tracker. The issues contain starting files, a first bounded increment, dependencies, acceptance criteria and pickup instructions. Read their current notes before claiming work; this map does not reassign an active owner.

Release boundary and accepted starting point

The bounded 0.18 controls are qualified at containing main bc67d176. Their benchmark records source/native identities, complete calculations, strict refusals, independent comparisons, cost and exclusions. Reuse that evidence for unchanged behavior. It does not establish whole-method automatic accuracy or production speedup. Candidate acceptance remains #557; the wider work below does not enlarge that release gate.

Existing molecular scalar/range/FT, direct/RI/BUPO consumers and finite-source periodic HF/KS implementations are the starting point, not work to recreate. An old stage label is not evidence that a whole row is either absent or accepted. Conditional matrix entries require an implemented same-source route or a documented mathematical obstruction and explicit refusal.

Implementation packages

Completed packages retain their source and acceptance links; read the current issue state before claiming work. Remaining packages stay independently claimable within their stated dependencies.

Issue

Design stage

Remaining deliverable

#618

P0/P1

role-specific capability query, planner-derived admission for the kernel, one-electron and kernel-free producers, and the limit inventory (deep signed contractions, ghost center) are on main; remaining provider capability and libint coexistence coverage continues

#619

P1c

libint-backend range-separated exchange (exact J/K), the shared guesses, spin seeds, ECPs (energies and exact/RI geometries), post-SCF dispersion, RHF/UHF harmonic frequencies, same-source atomisation references, QTAIM topology, molecular smearing and DFT+U energies and gradients are on main; remaining molecular method (ROHF/ROKS) and response compositions continue

#620

P2b

add bounded large-metric RI and RI-BUPO solves

#621

P2a/P2b

implement certified incremental direct and RI-BUPO builds

#622

P3b

stored-graph pair-tree metric bounds accepted at 27d9c24d; actual contraction arithmetic remains #624

#623

P1/P3b

bound integral and stored-source construction arithmetic

#624

P2/P3b

paired/retained Coulomb, evaluated assembly, direct K, BvK correction and whole-mesh fitted K accepted through 9e3a8c9a; input/XC and remaining arithmetic errors remain

#625

P3b

opt-in omitted-image overlap rank guard connects shared spectral bounds to fixed-source and support-selection admission; conditional metric assembly rank and finite inverse/projector response are implemented; full construction arithmetic and occupied-space stability remain

#626

P3b

signed range admission and grouped reference bounds accepted through da33a7dd; XC and useful whole-method bounds remain

#627

P3b

fixed-density certificate composition with no double counting, outward scalar summation, fail-closed bound admission and source/family-qualified full-Coulomb replacement and disjoint fitted-K composition for HF and signed global/range KS, plus an uncoupled second-order relaxation delimiter, are on main (component_ledger); relaxation and whole-method bounds, automatic selection remain

#628

P3b

extend direct periodic BUPO electrostatics to complete KS compositions

#629

P3a

qualify same-source BIPOLE primitive substitution

#630

P3c

derive and implement slab and wire source families

#631

P4

complete union-and-weight and WSSC provider adapters

#632

P4

complete neutral Bloch and real-Gamma MANTA source coverage

#633

P4

complete finite-character chi MANTA source adapters

#634

P4

resolve conditional semiempirical long-range reuse

#635

P5

qualify source-specific integral and BUPO symmetry reduction

#636

P6a

complete selected KS and BUPO molecular derivatives

#637

P6a

complete periodic and CCM gradients and cell stress

#638

P1b/P2/P3b/P6b

matched direct J/K baseline, shell-quartet reuse and the iterative R-tensor table, kernel gather tables and the shell-pair expansion cache measured (benchmark, ratio 22 -> 5.6); ket-side contraction rejected for high-l cancellation; RI versus RI-BUPO crossover measured through complete selected RHF on water chains (crossover benchmark): no crossover at n <= 8, RI-BUPO 3.0-3.9x the RI-J build with 75k-777k far blocks at energies equal to 3e-13 Ha, and 74% of the SCF in the far-block admission repeated every build; the order/leaf study (#757 step 1) takes RI-BUPO from the fixture’s 3-4x (20-35x with K screened) to 0.93x (n = 4) and 1.32x (n = 8) the RI-J build at leaf 8 / order 6 with the exact K screened, and the #757 step 2 verdict retention (admission walk recorded once, replayed per pass) to 0.85x and 0.68x at bitwise-equal energies, the first measured RI-BUPO crossover; #748 slice 3 then screens the RI-J three-center passes (RI-J build 1.3x / 2.2x faster, 574x / 135x libint direct at n = 4 / 8, energy bound 220-360x above the actual 1e-9 Ha error), which puts RI-J ahead of RI-BUPO again until its near tiles are screened (slice 3b); #750 slice 0 shares the fitting provider’s pair expansions (bit-identical, 1.06x on the n = 8 build) and slice 0b hoists the pair expansions inside the shared-channel kernel (bitwise A/B on 128 J/K matrices): screened RI-J build 2.0x faster at n = 4 and 1.8x at n = 8, 60x libint direct at n = 8, with the primitive Coulomb kernel now 52% of the SCF; #749 step 1 digests each symmetry-unique quartet once (A/B 6.3e-13): RI-J build 2.3x / 1.9x faster, 38x libint direct at n = 8, three-center passes now 71% of the SCF; step 2 evaluates each unique orbital pair once in the RI passes: 28x libint direct at n = 8, the primitive Coulomb kernel now 53% of the SCF; #752 grid-Taylor Boys (at most 1.7 ulp): 1.14-1.19x per build, 23x libint direct at n = 8; #751 step 1 ket-first contraction plus measuring at the route’s default channel batch 1 (the benchmark had used 2): 20x libint direct at n = 8; retaining the RI three-center coupling when it fits the budget (bitwise equal to streaming): 7.8x libint direct at n = 8, and the certified Schwarz screening of #748 cuts the exact K build 6.8x (n = 4) and 19.6x (n = 8) at tau = 1e-10 with actual errors 2500x inside the receipt bound, taking the ratio to libint direct from 5300-6300x to 790x and 320x; the unique-quartet digestion, block contraction and three-center screening follow in #749-#751

#639

P6b

qualify optional GPU, distributed and mixed-precision execution

Capability matrix coverage

Every row of the engine capability matrix maps to the issues below. A row can have several consumers or conditional source proofs; completing one consumer does not close the row. These are work assignments, not new claims of supported methods or accepted accuracy.

Kernel or consumer row

Implementation and acceptance issues

Molecular S/T/nuclear, cross-basis S and moments

#618; method composition #619

Two-center metric and three-center coupling

#618; bounded solves #620; arithmetic and rank #623-#625

Four-center Coulomb and direct J/K

#618, #619; incremental builds #621

erf/erfc two-, three- and four-center kernels

#618; periodic range accuracy #626; source adapters #629-#633

Molecular exact J and exact K

#618, #619, #616

Molecular RI-J and RI-K

#619, #620, #624; independent K admission

Molecular BUPO-J and RI-BUPO-J

#619-#621; measured crossover #638

Gamma and multi-k exact/RI J/K

#576, #607; accuracy #623-#627

Periodic BUPO-J and RI-BUPO-J

#622-#628; large metrics #620 and cost #638

Nuclear gradients and cell stress

#636, #637; separate derivative capability admission

Numerical source row

Implementation and acceptance issues

Molecule or isolated finite cluster

#618-#621, #616

Three-dimensional Gamma

#576, #607, #622-#628

Three-dimensional multi-k and shifted meshes

#576, #607, #622-#628

One- and two-dimensional periodic

#630

BIPOLE active physical or historical source

#629; shared domain/state owners #531 and #606

CCM union-and-weight and WSSC four-center

#631

Neutral CCM Bloch

#632

Neutral CCM real-Gamma

#632 and existing source integration #519

Finite-character chi four-center

#633

Finite-character chi RI

#633

Finite-character chi RIJCOSX

#633; preserve the existing COSX K source

Low-dimensional CCM

#630-#633; each named source needs its own proof

SECCM, MSINDO-CCM, PM6, OMx, DFTB and GFN2

#634; conditional long-range subterm reuse only

Source-qualified symmetry is #635 with shared admission #577/#608. Molecular and periodic derivatives remain #636/#637; CPU dispatch and accelerators remain #638/#639. Correlation tiles and factors reuse #581. These obligations apply across the relevant rows and do not follow from a passing scalar-kernel or Fock-action test. Conditional cells require the named same-source implementation or a documented mathematical obstruction and explicit refusal.

Existing issues reused

Issue

Responsibility

#616

Molecular evaluated-state integration alongside the libint bridge

#607

Periodic bounded native evaluated-state/solver adaptation

#576

Complete multi-k HF/DFT/hybrid MANTA control and parity ledger

#581

Molecular/periodic correlation tile and factor admission

#519

Neutral real-Gamma RG-8 source and consumer integration

#577

Cross-method symmetry and Fock parity

#567

Existing citation-route defect

#557

Bounded 0.18 candidate acceptance only

#643

Explicit automatic reference/metric and component reservations qualified at 4736aceb; original insufficient-cap refusals preserved

Shared Fock construction and eigensolvers remain under #472; PBC/BIPOLE and CRYSTAL23 parity under #593; full-SCF cost acceptance under #574; DLPNO under #508. These issues are dependencies and consumers, not competing implementations. MANTA provider symmetry uses the shared algebra and admission authority; no transport test alone permits production reduction. Correlation needs qualified integral tiles/factors, not merely a J action. Library/dependency defects retain their existing tracker owners.

Shared interface ownership

Interface

MANTA work

Shared implementation and acceptance

Molecular physical Fock and evaluated state

#619 and #616

#472 and #594; use the same borrowed/owned native bridge as libint

Periodic physical state and eigensolution

#607

#472; preserve bounded native execution, spin/k frames and physical versus solver-modified operators

Periodic sources and complete-method parity

#628, #629 and #630

#593 and #576; BIPOLE domain/cancellation policy stays with #531, its Fock adapter with #606

Source-qualified symmetry reduction

#635

#577 and #608; shared algebra does not grant producer or broken-state admission

Complete-method cost

#638

#574; reuse accepted source-matched controls and include setup, terminal work and peak memory

The operator pickup map and PBC roadmap link back to this MANTA tracker. When changing a shared producer or adapter, record the specific API, affected consumer issue and owned files on the work item. Link the landed source and applicable evidence from both sides of that dependency. Retain source/metric receipts separately from AO, spin, k-point and solver frames. Auxiliary-metric factorization is not an orbital Fock eigensolve. Coordinate the particular missing interface; do not wait for all parent issues to close or create another state object or review chain.

The concrete shared bridge #594, complete-method parity #576, BIPOLE state adapter #606 and cost acceptance #574 also link back to #617 and their respective MANTA consumers. Update the implementing issue and the affected consumer when an interface lands; reuse the same source and evidence links.

Production readiness ledger

The ordered speed programme (certified screening, block matrix-MD contraction, SIMD Boys, Split-RI-J, periodic generalization, exchange, BUPO regime, hybrid dispatch and the paper package) is the performance plan with tracker issues #747-#758, created 2026-10-30.

Snapshot of 2026-10-24 at containing main 4736aceb, reconciling every capability-matrix row with what is implemented, what carries recorded numerical acceptance, what the manual supports within an explicit envelope and what stays experimental. Closed-issue counts are not readiness; an E or stage label in the matrix does not qualify its row. Four states are used. Implemented: the path exists behind its method guards and its focused tests pass; no independent numerical acceptance is claimed. Validated: a complete affected-method calculation was compared with an independent reference (ORCA/CRYSTAL, an independent NumPy/SciPy or Gaussian/Boys/Fourier construction, or a known limit) with recorded source/native identity, inputs, tolerances and measured errors, within the limits those records state. Supported: validated, documented with its envelope (methods, sources, bases, limits, cost), refused outside it before expensive work, and no longer labelled experimental for that envelope. Experimental: everything else, including validated components whose whole-method accuracy, cost or failure behaviour is not established. Today no row reaches Supported: every MANTA route keeps the experimental label (manual, engine guide, release controls).

Baseline refreshed on 2026-10-24

The bounded 0.18 MANTA contribution (#557) is a PASS for candidate 5dfc97f0: build 24463 and candidate pipeline 14816 succeeded, the final native receipt’s core SHA256 and its 17 dependency hashes match the verified predecessor, and the receipt’s native tree equals the candidate tree (tracker #557, 2026-10-24). The accepted four-profile evidence remains the local calculation set vq9c085769c600 on its own local core: 8 converged calculations, 10 strict physical refusals and 1 expected old-cap refusal, reusing the original HF and continuous-grid PBE0 component and independent references (benchmark). The local scientific core and the Linux candidate core are distinct binaries; neither is a deployed runtime, and no cross-platform numerical parity is claimed. That acceptance covers the four frozen profiles only (shifted two-point Be RHF/PBE0 and a compact Gamma He RHF with RI-J, exact full-Coulomb K, fixed supports, unreduced meshes). On 2026-10-24 the compact-tight profile re-ran bitwise unchanged on main d8f937fb after the #638 kernel change (all 1e-8 targets met; exchange 8.9e-12 / 3.0e-10 Eh, nuclear 1.9e-17 Eh, kinetic 3.9e-13 Eh; local core, not a candidate binary).

Kernel and consumer rows

Row

State

Evidence and envelope

Open

Molecular S/T/nuclear, cross-basis S, moments

Implemented; Bloch S/T validated

One-electron operations through l=6 and moments to order 6 behind one_electron_capability; ECP stays its own provider. Periodic Bloch S/T parity ≤4.5e-15 (benchmark); no hashed molecular acceptance page

#618

2c metric and 3c coupling

Implemented

FittingTiles auxiliary l≤7, orbital l≤6, full/erf/erfc, no derivative; dense metric/factor only

#620 bounded solves; #623-#625 arithmetic and rank

4c Coulomb and direct J/K

Implemented; consistency-checked

Direct digestion through l=6 matches the libint reference to 1e-10 Ha (water/cc-pVDZ) and to 3e-12 Ha on four small fixtures with identical shell schedules; 9.5x slower than libint after shell-quartet reuse (cost benchmark); certified Schwarz screening opt-in through screening_tau (#748 slice 1: omitted quartets, max-norm J/K bounds and the energy bound in the receipt; default 0 is bit-identical; exact K 6.8x/19.6x faster at n = 4/8; slice 3 extends it to the RI-J three-center passes with a metric-weighted J bound and an energy bound on the crossover page)

#638 kernel arithmetic and dispatch; #621 incremental builds

erf/erfc 2c/3c/4c kernels

Implemented; periodic component bounds validated but loose

Kernels exist; the erfc quadrature target 2e-14 is not a rigorous bound; the libint backend refuses range-separated exchange. Periodic raw and pre-SCF signed range bounds accepted at f1698506 and 1c06da0c with matrix bounds 55-247 and energy bounds 1.8e4-8e4 Eh (metric benchmark)

#626 useful bounds; #629-#633 source adapters

Molecular exact J and K

Implemented; consistency-checked

RHF/UHF/RKS/UKS gas phase through run_* and run_job with the shared guesses, spin seeds, ECPs, dispersion, atomisation, QTAIM, smearing, DFT+U energies and gradients, HF gradients, optimisation and RHF/UHF frequencies; refused: ROHF/ROKS, response, correlation, solvation, RKS/UKS and DFT+U Hessians. No hashed acceptance page and no matched ORCA comparison recorded

#619 remaining compositions; #616 evaluated-state bridge

Molecular RI-J and RI-K

Implemented; consistency-checked

All eight J×K selections; RI HF Lagrangian gradients including fitted ECP results; RI-J −29 µHa and RI-J/K +21 µHa versus exact on water/cc-pVDZ (manual, unhashed)

#619, #620, #624

Molecular BUPO-J and RI-BUPO-J

Implemented; construction validated

Exact terminal confirmation; J error 8.6e-14 and energy 1.3e-13 on the benchmark cases, but no far block passes strict admission on water or the dimers and optimized libint is faster than every MANTA path (BUPO benchmark); through complete selected RHF on 10 bohr water chains RI-BUPO-J agrees with RI-J to 2e-13 Ha but costs 3.0-3.9x per build at n = 4-8 with the far-block admission dominant (crossover benchmark); the #757 order/leaf study and verdict retention bring RI-BUPO to 0.85x (n = 4) and 0.68x (n = 8) the RI-J build at leaf 8 / order 6 with the exact K screened (energies bitwise equal, bound 2e-7 to 9e-7 Ha) against the fixture’s 20-35x before retention; the admission walk now runs once per geometry and the far field’s remaining cost is the per-pass matrix formation; gradients and incremental builds gated

#620, #621, #757 retention and order ladder

Gamma and multi-k exact/RI J/K

Implemented; component-validated on frozen profiles

Neutral all-electron 3-D cells with explicit fixed supports, HCORE/AUTO guess, no DIIS; release-control profiles accepted (above); fitted-K radii ≤1.9e-15 Fock and ≤9.4e-14 Eh accepted at 9e3a8c9a; PySCF AFT parity 3.5e-13 Eh and complete RHF reference 1e-15 Eh (unhashed pages). No automatic accuracy profile: domain series versus GDF reach +0.74 mHa (Gamma) and −0.15 mHa (2,2,2) only at the largest domains

#576 complete parity; #607; #623-#627

Periodic BUPO-J and RI-BUPO-J

Implemented (HF, exact reciprocal K); experimental RI multipoles

Public RHF/UHF only; RI multipoles labelled experimental, qualified for RI-J with exact K or pure KS; BUPO J ≤2.0e-13 / energy ≤2.8e-14 on sources b6054324, 4c5ceb97; public job F ≤2.6e-13 / energy ≤1.9e-11 on 06731fc7; RI accuracy page 82547576: actual errors ≤1.3e-16 against bounds up to 8.8e-6; pair-tree metric bounds at 27d9c24d; complete RHF slower than exact finite RI

#622-#628, #620, #638

Nuclear gradients and cell stress

Implemented (molecular HF exact/RI incl. ECP, DFT+U); absent periodic

Molecular HF gradients validated against finite differences and the legacy gradient in tests, no hashed page; KS gradients, BUPO gradients, periodic gradients and stress refused

#636, #637

Numerical-source rows

Source

State

Evidence and envelope

Open

Molecule / finite cluster

Implemented

Rows 1-7 above; no hashed acceptance, experimental label

#616, #618-#621

3-D Gamma; 3-D multi-k and shifted meshes

Component-validated on frozen profiles

Release controls on 7917b560, grid integration 9a868249, corrected candidate 2d0fdb8c and today’s d8f937fb re-run; image 1e-6 and compact 1e-8 targets met; image-tight RI J 1.3e-9 / energy 2.5e-7; PBE0 versus independent reference 5.3e-15 Eh. Excludes whole-method relaxed energy, basis/k/XC convergence, bulk parity, symmetry and speedup

#576, #607, #622-#628

1-D / 2-D periodic

Not implemented

Periodic route is 3-D only; dimensionalities refused; low-dimensional boundaries excluded from the release scope

#630

BIPOLE active/historical source

Not implemented

No MANTA adapter; supports do not substitute BIPOLE AC/BD domains

#629 (#531, #606 owners)

CCM union-and-weight / WSSC

Not implemented

CCM MANTA routes unqualified

#631

CCM neutral Bloch

Not implemented

Same

#632

CCM real-Gamma

Partial

MANTAFitSource, MANTAOneElectronSource, MANTAXCSource fold sources reproduce fitted HF/KS energies per cell under the experimental AICCM route; no hashed acceptance

#632, #519

chi four_center / ri / rijcosx

Not implemented

No adapter documented

#633

Low-dimensional CCM

Not implemented

#630-#633

SECCM / MSINDO-CCM / PM6 / OMx / DFTB / GFN2

Not implemented

Conditional long-range reuse only

#634

Cross-cutting

Symmetry: the finite-product algebra correction d5a519cb is an experimental handoff, not producer qualification; the bounded release controls do not depend on it (#635). Arithmetic: paired, evaluated-assembly and fitted-K arithmetic are accepted through 9e3a8c9a; input/XC construction and remaining non-paired arithmetic stay open (#623, #624). Automatic accuracy: the three fixed-density certificates are separate ledgers with different reference objects and must not be summed into a relaxed-energy tolerance; no relaxation, overlap-rank or k-mesh contribution exists (#625, #627). Cost: MANTA is slower than libint on every measured path (320-790x on screened water chains after #748 slice 1, from 5300-6300x unscreened) and RI-BUPO at leaf 8 / order 6 is 0.85x (n = 4) and 0.68x (n = 8) the screened RI-J build after the verdict retention (#638, #757 steps 1-2); with the RI-J passes screened too (#748 slice 3) RI-J leads again, 135x libint direct at n = 8, until the RI-BUPO near tiles are screened. GPU, distributed and mixed precision: not started (#639).

What this means for a production verdict

Supported configurations: none without the experimental label. The narrowest candidate envelope is molecular RHF/UHF/RKS/UKS with exact or RI J and K on all-electron and ECP Hamiltonians, where results match the libint route to 1e-10 Ha and gradients match finite differences, but it lacks a hashed acceptance page, a matched ORCA comparison and acceptable cost. The periodic envelope is the four frozen profiles with fixed supports, which carry component certificates but no whole-method accuracy or bulk parity. Failure behaviour is explicit: unsupported compositions refuse by name before the basis is built, and unmet accuracy targets refuse before an archive is written. Remaining restrictions are the open rows above; the release-owned 0.18 scope stays unchanged by this ledger.

Parallel work and acceptance

Useful independent starts include the capability inventory, molecular job composition, same-source BIPOLE adapter, conditional source proofs, derivative primitives and benchmarks using already accepted APIs. Pair-tree stored-graph metric bounds and the executed paired metric substep, including its exact-adjoint J contribution, individual coupling actions and their paired J/energy composition, including spin/k weighting, evaluated HF/KS assembly and stored-input direct K, BvK correction, retained J and whole-mesh fitted K, including signed evaluated Fock/energy composition, are accepted. Input/XC construction, remaining non-paired arithmetic and automatic periodic accuracy remain the ongoing MANTA integration lane. The paired metric evidence is recorded in the metric benchmark. It does not close #624 or certify a complete arithmetic budget. The exchange fitting benchmark qualifies a separate full-Coulomb K component bound under #626. Its uniform family bounds remain conservative. RI HF attachment is accepted at 2dfbd54d; Global-hybrid KS attachment is accepted at 2dfb076d; raw erf/erfc component bounds are accepted at f1698506; pre-SCF signed range composition is accepted at 1c06da0c. Optional grouped reference bounds are accepted at da33a7dd. XC, further reference tightening and useful source selection remain #626/#627. Claim a bounded scope and intended files on the selected issue before editing shared producers, and preserve existing owners, jobs and checkout changes. Unassigned does not mean unowned. A dependency is the specific needed capability, not a wait for every umbrella issue to close.

Use existing producer, Fock, numerical and symmetry APIs. Keep physical source, normalization, support, rank, zero-mode, charge and evaluated-state identities explicit. No silent dense fallback, source substitution, default change or removal of experimental labels follows from this tracker.

The implementing chat may implement, validate and accept its own increment. There is no mandatory second reviewer or review-only hold. Land focused, checked source promptly with ordinary CI skipped. Scientific acceptance and issue closure require a complete affected-method calculation through vq, matched to ORCA/CRYSTAL where suitable or to justified numerical/high-precision references with limitations stated. Reuse valid evidence and a healthy runtime containing the fix; record actual source/native identity, inputs, job/output and comparison. Full CI is reserved for release candidates. Update the manual, benchmark, citations and changelog when corresponding behavior ships.

Keep #617 open until every original matrix and stage obligation is accounted for by accepted work or an expressly conditional obstruction/refusal. Finishing #557 or one bounded work package does not complete the full goal.

The optional stored-operand fitting metric assembly guard advances #623/#625: it feeds actual tile-source entry enclosures into the existing strict spectral rank radius. Primitive/Fourier/phase/weight construction errors and inverse response remain separate open work. It does not retire the full source-error, automatic-accuracy or production-readiness obligations under #617. See scientific contract Section 5.55 and the metric spectrum benchmark.

Finite retained-projector and truncated-inverse response now consumes the conditional #623 assembly perturbation in the shared inverse certificate (#625, contract 5.56). It includes partial-rank rotation and moving relative cutoffs, and composes the stored-factor error without changing factors. Coupling/source construction and J/K/energy propagation remain open; this adds no whole-method accuracy or production verdict.

The optional #627 metric-action target can now include the #623/#625 conditional metric-assembly inverse response, propagated through the shared explicit or supported compressed coupling norms and fixed density support. Separate source/stored components avoid double counting; default physical sources and the global ledger are unchanged. Primitive/coupling construction, K propagation and total relaxed accuracy remain open (contract 5.57).

Conditional metric-source K response (#626) now has separate optional raw component targets across actual Gamma and ordered q owners, including physical range kernels (contract 5.58). It consumes the shared #623/#625 assembly/inverse proof and coupling norm producer. Existing fitting/signed aggregate/global ledgers stay distinct. Source operands, coupling, metric domain/screening response and relaxed accuracy remain unfinished under #617.

The #623 short-range metric Bloch phase is now enclosed from exact stored q/R vectors, including dot-product and sine/cosine evaluation, by the shared bounded Taylor/angle-doubling proof (contract 5.59). Existing optional metric assembly/rank and J/K source-response targets consume it. Physical sources and states remain unchanged. Primitive/Fourier/coupling-phase/weight and q/R construction, domains and whole-method accuracy remain unfinished.

The #623 reciprocal metric-weight construction proof (contract 5.60) now includes stored-vector squares, stored range parameters and exp evaluation, using the shared interval producer and existing physical kernel/zero modes. Metric rank/inverse and conditional J/K targets consume the expanded proof. The reciprocal prefactor, Fourier/coupling construction, domains, relaxation and useful automatic whole-method accuracy remain unfinished under #617.

The #623 reciprocal metric prefactor proof now includes pi and the exact stored-lattice determinant (contract 5.61), consumed by existing weight, rank/inverse and conditional J/K producers. Strict unresolved-volume refusal preserves the physical cell. Primitive/Fourier/coupling and vector generation, domains, relaxation and useful automatic whole-method accuracy remain open.

The #623 auxiliary Fourier construction enclosure (contract 5.62) now covers primitive Gaussian moments/decay, phase, signed contraction and arithmetic on stored representation coefficients across auxiliary l=0..7. The existing metric/rank/inverse/J/K chain consumes it; numerical panels remain unchanged. Basis/representation data generation, short-range/coupling construction, vector generation, domains and useful relaxed accuracy remain open.

The #623 two-center metric construction proof (contract 5.63) encloses radial seeds, primitive prefactors, center shifts, Hermite contraction and application of stored pure transforms through l=7. It feeds shared rank/inverse/J/raw-K bounds without changing numerical source or states. Basis/representation and vector-data generation, three-center/AO-pair coupling construction, domains, screening and relaxation remain unfinished; #623/#625/#617 stay experimental.

The #623 positive erfc radial proof (contract 5.64) tightens the short-range construction component without changing its reference or numerical producer. Bernstein positive-series panels and explicit tails avoid full/long-range subtraction; the ordinary Boys proof remains shared and unchanged. Existing metric/rank/inverse/J/raw-K consumers propagate the result. Useful automatic accuracy, coupling/data construction, domains and relaxation remain open.

The #623 representation construction proof (contract 5.65) removes the stored pure-matrix premise from auxiliary Fourier and two-center metric certificates. Exact rational-polynomial/sqrt bounds feed shared rank/inverse/J/raw-K without changing numerical transforms or states. Basis/vector data generation, coupling, domains/screening and relaxed response remain open; no symmetry permission is inferred. #623/#625/#617 remain experimental and unfinished.

The #623 orbital-pair Fourier construction component (contract 5.66) adds a shared certificate through orbital l=6 for Gaussian products, exact product centers, phase/moments, contractions and both pure transforms. It retains the numerical source and is currently a separate diagnostic. Full three-center coupling composition and propagation, stored input generation, domain/screening and relaxed response remain open. No existing metric/Fock/symmetry admission inherits this component automatically; #623/#617 remain experimental.

The #623 streamed three-center construction component (contract 5.67) covers the exact stored auxiliary/orbital Gaussian integral through l=7/6/6 for Coulomb/erf/erfc. Shared radial, elementary and representation proofs feed per-tile Hermite/Coulomb construction bounds without changing numerical tiles. Full periodic coupling composition and action propagation remain open, along with stored-input/image generation, domains/screening and relaxed accuracy. This component does not expand existing metric/Fock or symmetry admissions.

The #623 finite periodic coupling diagnostic (contract 5.68) composes shared three-center, Fourier, phase, lattice-weight and representation proofs over literal selected integer image sets, including translated centers and the finite-zero term. Separate tile/forward/adjoint certificates compare unchanged numerical routes directly with that unscreened finite reference. It closes the finite construction composition gap; propagation into J/K/RI-BUPO/SCF admission, input/domain completeness and relaxed accuracy remain open. Existing Fock/PBC/symmetry interfaces and metric budgets are preserved. #623/#617 remain experimental and unfinished; overlapping component budgets must not be added.

The retained-source discrepancy #687 is repaired by sharing the existing budgeted primitive selector between MANTA and libint, including dense and streamed paths. Screened source identities and split-call admission reflect the retained primitive set; unscreened sources stay unchanged (contract 5.69). Finite coupling/metric composition into J/K/RI-BUPO/SCF remains the next open accuracy increment; the primitive repair does not close #617 or that work.

The accepted #627 producer seam combines finite coupling construction, stored factor and metric-source response into a declared fixed-density J/energy certificate on the actual Gamma RI owner (contract 5.70). Explicit and compressed/global/pair-tree producers share the same reference and keep separate component radii. Complete-job automatic admission and reference- matched BUPO ledger composition, K response, metric translation generation and relaxed accuracy remain open; this standalone joint certificate does not close #627 or #617.

Local complete-method controls cover six explicit HF/KS profiles and three compressed/global/pair-tree profiles, with the existing independent references reused. The joint bounds still do not meet useful automatic accuracy targets. See the finite-source response evidence for component bounds, actual cost and the source/runtime qualification boundary.

The accepted #627 integration applies that joint certificate as a separate optional pre-SCF finite-source target in the existing complete-method selection loops (contract 5.71). Shared count planners reserve only requested construction proofs and retained trial receipts. This does not combine unlike full-Coulomb references, improve the bound magnitudes, or close the remaining source/domain, K and relaxed-state accuracy work.

The bounded admission passes nine unchanged complete-method profiles and selection/refusal controls. It does not establish useful automatic accuracy: a larger auxiliary control still has a very large finite metric-response bound. Preserve that failure of useful accuracy while reusing the accepted admission/state evidence; see the complete-method admission benchmark.

The accepted #627 residual metric and coupling responses (contracts 5.72 and 5.73) tighten the joint finite-source certificate without changing the physical source. The compact three-to-eight auxiliary RHF profile now passes finite J/energy targets 3e-8/1.5e-6 Eh and refuses an energy target of 1e-6 Eh within its existing search limits. The measured joint energy bound is 1.1078796248453086e-6 Eh; this is a fixed weighted-family finite fitted J response, not a total relaxed-energy tolerance. Compressed norm fallbacks and all other source/method matrix rows retain their existing qualification. See the finite coupling response evidence.

The next accuracy work must address the explicit remaining requirements: metric translation/input generation and reference-aligned composition with full-Coulomb fitting and executed arithmetic, followed by missing K/XC and relaxation contributions. Do not sum unlike ledgers or infer generic diffuse/ skew-cell, independent-spin, range, provider, derivative, GPU/distributed or symmetry readiness from this one bounded profile. Full #617 stays open.

Exact finite metric lattice images

The shared metric proof now encloses exact lattice-times-integer image centers and phases, matching the coupling convention for the selected integer set. Implementation140b95e0/reference repair506e14c7 is accepted through local vq 52f189aae41f: three unchanged complete states and three strict refusals. The larger eight-function control still admits finite J/E targets3e-8/1.5e-6 and refuses energy1e-6. Metric assembly/joint reference identities advance; source, rank, density and Fock identities remain unchanged. See the acceptance and exclusions.

This closes only selected metric image generation. Input/reciprocal generation, image-domain completeness, infinite tails, full-Coulomb fitting, K, executed Fock/energy arithmetic and nonlinear relaxation remain separate. No total relaxed-energy, bulk CRYSTAL parity or speedup claim follows. Reference-aligned whole-method composition and the rest of #617 remain open and experimental.

Evaluated finite Coulomb response

Source1dee3252 composes the immutable finite-source certificate with one retained or paired executed-arithmetic proof at the evaluated stored spin density and weights. Local vq922d68c30af0 passes four unchanged complete states and three strict refusals. The eight-function auxiliary evaluated Coulomb energy bound is1.1078799401486478e-6 Eh. Existing pre-SCF targets remain source-only; the evaluated receipt does not silently extend their semantics. See bounds, costs and independent-reference limits.

This closes the raw J/Coulomb response arithmetic composition for the admitted finite reference. Restricted embedding, final assembly, physical/reference completion and nonlinear relaxation remain separate. The unlike full-Coulomb ledger is unchanged. Full#617 and useful automatic whole-method accuracy stay open and experimental; no total relaxed-energy, parity or speedup claim.

Euclidean finite-source metric envelopes

The shared owner-sealed column proof now also encloses Euclidean norms using scaled outward sums of squares (contract 5.78). Residual, inverse action and complex transpose/adjoint projection bounds retain the old envelopes as fallbacks. This reduces auxiliary-dimension loss without changing the numerical source, retained rank, fixed density family or shared Fock/PBC/ symmetry semantics. The existing L1 coupling norm keeps its interpretation. See the bounded acceptance.

The full #625/#627/#617 scope remains open. Source-only targets do not certify nonlinear SCF relaxation, and the finite-source reference is not the infinite full-Coulomb reference. Input, domain, exchange, assembly and relaxation gaps remain explicit; this increment grants no new symmetry or performance claim.

Finite Coulomb through evaluated assembly

The shared assembly audit now consumes the immutable finite Coulomb response and binds it to actual density/weights, held inputs and evaluated witnesses (contract 5.79). It adds assembly-only arithmetic once, closing final Fock and energy assembly for the exact finite Coulomb reference with other supplied components held fixed. Existing combined arithmetic and full-Coulomb ledgers remain separate. See the assembly acceptance.

This does not certify construction or response of the held non-Coulomb components, ideal restricted embedding, physical-domain completion or SCF relaxation. Pre-SCF source-only targets and Fock/PBC/symmetry ownership are unchanged. The full #624/#627/#617 scope remains experimental and open.

Coulomb density embedding at the physical input

The shared finite source certificate now exports reference-column/inverse envelopes to the existing assembly audit (contract 5.80). Its separate finite Coulomb reference uses physical stored density, including restricted half-spin embedding loss. Exact conversion adds zero; subnormal loss retains the linear J and quadratic energy terms. See the embedding acceptance.

Larger eight-auxiliary RHF, tree UHF and continuous-grid PBE0 RKS retain accepted states bitwise. Prior tighter metric bounds remain unchanged. Source-only admission, explicit compression versus auxiliary selection and shared Fock/PBC/symmetry contracts are preserved. Other physical components remain held fixed; non-Coulomb embedding/construction, domain completion and density relaxation are still excluded. The full #624/#627/#617 scope remains open and experimental, with no total relaxed-energy or general performance claim.

Euclidean coupling-construction error columns

Finite-source coupling construction and compressed replay radii now use the shared scaled Euclidean aggregator with the original L1 fallback (contract 5.81). This reduces auxiliary-dimension loss without changing the stored source, rank, density family, Fock/PBC/symmetry contracts or exact finite reference. See the component acceptance.

Three complete states remain bitwise unchanged. The eight-auxiliary finite Coulomb plus assembly energy bound tightens from 7.39184e-7 to 7.14015e-7 Eh; metric-only and stored-factor components remain unchanged. Targets retain source-only semantics, and explicit compression remains separate from automatic auxiliary selection. Physical-domain completion, other-component errors and relaxation remain excluded. Full #625/#627/#617 remain open and experimental; no whole-method accuracy or speedup claim is added.

Terminal finite Coulomb assembly targets

Optional paired terminal Fock/energy targets now consume the shared finite Coulomb assembly receipt for complete RI HF/KS (contract 5.82). These are explicitly delimited non-relaxed targets with other supplied components held fixed. Pre-SCF source targets/search, source/density/rank and Fock/PBC/symmetry contracts remain separate and unchanged. See the terminal-target acceptance.

Four converged states match accepted evidence bitwise. The eight-auxiliary RHF case meets its 7.2e-7 Eh terminal finite-reference energy target. A stopped state can pass its target while remaining unconverged; an impossible terminal target refuses without a qualified QVF. Physical-domain completion, other component errors and relaxation remain open. Full #627/#617 remain experimental and unfinished; no total relaxed-energy accuracy or speedup claim is added.