BIPOLE erfc continuation after the repository split

The physical quartet/nuclear-domain integration is recovered and remains opt-in. Terminal XC reuse, Schwarz self-norm screening and screened derivative buffer handling are corrected. LiH and Si native SCFs and fixed-density grid/cutoff checks are complete. The 840-case run completed with 836 passes, two diagnosed grid-fixture failures, one known expected failure and one historical skip. Both reviewed fixtures then passed their reruns, and the historically skipped MgO force comparison passed separately. Production code was unchanged between these runs. The global default and issue closure remain subject to the acceptance limits below.

Recovery, 2026-09-08

The continuation uses the standalone vibe-qc/vibe-qc repository, branch codex/bipole-erfc-resume, starting from bdec7d2. The initial pull was followed by the development install’s main synchronization. The preserved candidate was based on monorepo commit 5d49d0c28573.

The 50 surviving source, documentation and test files were restored from codex-bipole-full-wip-paused-20260908.patch under ~/vqruns/bipole-erfc-20260906/physical-coupled-wip/. The old handover directory was removed in the split; the archived patch retains its history. The suite manifest was regenerated from this repository’s current inventory instead of restoring the monorepo’s generated manifest.

Historical issue numbers in this note refer to the monorepo. The outstanding work was tracked as #21, #66, #674, #724 and #20. Their new repository IDs are #186, #182, #82, #44 and #187, respectively; see the issue mapping. At the user’s handoff, cluster deployment and bugctl SHA validation still targeted the monorepo. Their migration was not exercised by this local work; these results do not establish deployment or independent issue closure.

Operator and acceptance boundaries

The recovered physical-domain implementation is opt-in through pair_complete_1e. It coordinates AO-pair separation, quartet midpoint interaction reach, wider exchange output support, nuclear-source images, reciprocal support and their derivatives. Independent Gaussian sums, image relabelling, density adjoints and finite differences exercise the retained finite operator. Convergence of that finite operator toward the periodic limit is a separate requirement.

For a retained integral (ab|cd), both physical product separations must be within the pair cutoff, and the separation of (A+B)/2 and (C+D)/2 must be within the interaction cutoff. Coulomb and exchange apply this predicate to their respective integral pairings. In exchange, the external matrix indices need not themselves form a retained product pair. Restricting that external pair to the overlap support would truncate a different operator. The union of pair-centered spheres with radius equal to pair cutoff plus interaction cutoff encloses the required images without growing merely because an atom is given a distant lattice-image label.

Nuclear source selection uses a sphere about each AO-product midpoint. Reciprocal nuclear and Hartree terms use the same AO-pair support as their real-space partners. Derivatives preserve that support away from sharp cutoff crossings. Tests at a finite cutoff therefore establish the derivative of that retained sum, not uniform convergence across a cutoff boundary.

The global default remains off. Periodic COSX, diagnostic grid Hartree and explicit legacy Fock projection refuse the unsupported physical-domain combination. The #66 fixture retirement follows the prior maintainer decision to withdraw asymmetric legacy-HF converged-SCF claims; it does not loosen terminal stationarity tolerances.

The old native LDA finite-domain energy pin was reviewed separately from the independent periodic two-Gaussian LDA reference. Restoring only the historical grid partition for that exact fixture reproduces the old historical-domain and physical-domain values to 1.3e-13 and 1.4e-10 Ha. This establishes why the native regression anchors must follow the grid correction; it does not replace any independent physical reference.

On this build the historical-domain LDA result is -34.66112304044817 Ha (+0.25327789 mHa from its old pin); physical-domain LDA is -33.875342453487875 Ha (+0.25775326 mHa). These are measured discrepancies, now reviewed native regression anchors. Both independent HF split values and all four RKS/UKS periodic LDA reference witnesses passed in the domain run.

Validation record

The initially recovered extension’s SHA256 is fcc50e6eacf0d4034865095021bdfa47b298021e9a135eec2db1d6b767efd991. All 568 recorded native build inputs were unchanged when validation started. The package and extension both resolve inside this checkout.

Two binding checks passed. Four new RKS/UKS terminal-XC tests failed before the reuse change, each at the redundant call after the exact final-density phase. The complete terminal/helper/output selection then passed all 181 tests after the change, including converged and iteration-capped returns. The tests allow XC during a required final build and reject only subsequent quadrature. The returned XC energy must equal the exact build’s stored value. This three-file change is saved separately as local commit 3792411. The recovered physical-domain candidate was validated before committing the integration.

The full 424-case domain/consumer run completed in 906.63 seconds with 422 passes and two failures. Besides the preserved LDA pin, a runner test invoked a queue test file removed in the split. Replacing that external dependency with a synthetic selected pytest node preserves subprocess isolation and caller-state checks; all 48 gate tests pass after that repair. That repair is saved independently as local commit cdc6790. The original 424-case result remains recorded as a failed run.

Before correcting Schwarz self norms, at the saved LiH/PBE reference density, physical domains with a 20-bohr pair cutoff, 35.47-bohr resolved SR interaction reach and sr_image_precision=1e-8 give -8.103457093990926 Ha. Relative to the level-9 independent reference, the differences are +23.54956 microhartree total, -10.46735 microhartree Coulomb and +34.01692 microhartree XC. Kinetic and nuclear-attraction differences are at most 2.2e-11 Ha. This one-density evaluation took 515.92 seconds with one native worker; it is neither a converged SCF result nor a matched performance benchmark. Grid and screening ladders were then needed to explain the residual.

Tightening the native Schwarz threshold from 1e-12 to 1e-14 changes this Coulomb component by only -3.41e-10 Ha. An independent PySCF 2.14.0 fixed-density RSJK calculation reproduces the archived Coulomb reference at omega 0.4 and 0.7 within 5e-12 Ha. The native values at those split parameters differ by 7.18512 microhartree, which motivated the operator-domain/split investigation below. Independent long-range-only AFT calculations localize the discrepancy to the short-range erfc sum: after removing the common analytic background, long-range errors are -4.2e-14 and -9.65e-11 Ha at omega 0.4 and 0.7. The corresponding short-range errors are -10.46769 and -3.28248 microhartree. With charge-pair screening enabled at 1e-14, extending the physical interaction radius from 35.47 to 45 bohr leaves the Coulomb result bit-for-bit unchanged. The explicit-radius path does not automatically enable charge-pair screening. The home-cell J energy contraction agrees between charge screening on and off to about 1.4e-12 Ha. The full unattenuated-Schwarz 45-bohr run was stopped after these more targeted checks; its last row is incomplete and is not a convergence result.

Schwarz squared-norm correction

The residual comes from primitive screening in the Schwarz pre-pass itself. Libint’s default absolute precision is machine epsilon. A self integral below that precision can be discarded even though its square root permits mixed integrals many orders larger than the requested production cutoff. Reducing the caller’s Schwarz threshold cannot recover a bound already stored as zero.

Local commit da0534a disables primitive screening only while computing the molecular and periodic Schwarz self norms. The production ERI threshold is unchanged. Fourteen closed-form Gaussian regressions fail before this change and pass afterward, covering direct, molecular-limit and lattice J/K builders, bare Coulomb, two erfc kernels and two thresholds. Seven molecular incremental-Fock regressions also pass. The original three periodic screening checks pass, including their performance check.

At the same LiH reference density, the corrected Coulomb energies are 1.995852262476314 and 1.995852262374664 Ha at omega 0.4 and 0.7. They agree with the independent values to 1.7e-12 and 1.1e-10 Ha, respectively. This resolves the previously observed 10.47-microhartree Coulomb residual. The isolated first-build timings are 152.13 and 93.63 seconds versus 124.63 and 77.61 seconds before the fix; these observations are not a controlled performance benchmark.

The intermediate Schwarz-corrected extension’s SHA256 is de0e5158dfdf5e6eefc3a909c8d38315c5fa6a13cc1243fa9005ddbebe777a81. Only cpp/src/schwarz.cpp differs among the 568 native source inputs. The older gradient run was stopped as superseded and remains incomplete. The 816-test domain, gradient, convergence, terminal, output and screening selection exposed the derivative-buffer defect below and was superseded. Its completed shard and partial receipts are retained as historical results.

The native 15-bohr HF regression anchors move from -23.19211691121152 to -23.19210137102612 Ha for physical domains and from -22.93998258398289 to -22.939983004503866 Ha for historical domains. The original binary reproduces both old anchors within 8e-14 Ha; the native source audit isolates the change to Schwarz self norms. The image-relabeling invariant passes. The anchors have been updated without changing tolerances or independent physical references.

Screened derivative-buffer correction

Libint marks a fully screened two-electron quartet by clearing only the first result pointer. The periodic gradient contractions previously inspected each of the twelve derivative pointers independently, allowing stale pointers from an earlier quartet to contribute. Tightening force screening then made the error worse by admitting more calls that Libint subsequently screened.

Local commit 74cdf04 checks the whole-quartet sentinel independently for J and K before consuming derivative buffers. Eight of twelve new analytic Gaussian force cases fail before the change; all twelve pass after it. The MgO short-range J witness improves from a 1.34684e-6 discrepancy to 3.01438e-12 Ha/bohr against finite differences, with force balance restored. The complete periodic screening file passes all 29 cases in the isolated corrected build. Neither finite-difference step sizes nor tolerances were loosened.

The final production extension SHA256 is 853dca4cc0aa2d16238bdd70b7c2c35f86447cc8d90b67f1e4971376c703876b. Relative to the original native build inputs, only cpp/src/schwarz.cpp and cpp/src/periodic_gradient.cpp changed. A fresh complete 840-case selection includes the symmetry file and new derivative regressions. Four local shards use 1/1/2/1 native threads and one BLAS thread each, with exact collection inventory, source hashes and per-test receipts. Under shared-host load, the original two-hour monitor was replaced with another bounded two-hour monitor retaining the 16-GiB limit. The test coordinator and workers continued without restarting. The deliberate termination of the original monitor is recorded separately from pytest outcomes in verified/gate-guard-handoff.json. That attempt subsequently exceeded its aggregate memory cap: the monitor observed 17,425,219,584 bytes against a 16-GiB limit and stopped it. Its 614 passing calls and one strict expected failure are incomplete receipts; no test assertion failed. memory-limited-gate-stop.json records the stop.

The full 840-case selection is therefore restarted from scratch under verified-bounded/, with at most two concurrent workers, four native threads each, a 24-GiB aggregate limit and a four-hour wall limit. This run records every setup/call/teardown phase and hashes the Python production sources as well as the native inputs and selected tests. The source and extension are unchanged; only execution concurrency and resource limits have changed. This run exposed two further grid-related test assertions. The legacy RKS spheropole fixture’s old native energy pin is -1.1156375078921 Ha; the current partition gives -1.1156334486137565 Ha. Replacing only its grid builder with the historical partition gives -1.115637508045234 Ha, within 1.6e-10 Ha of the old pin. The spheropole is identical in both calculations. The recovered and final native binaries also give identical current-grid energies. legacy-rks-pin-review.json retains this isolation.

The LiH moving-grid XC witness reached its old minimum-correction assertion with a correction of 1.27478e-4 Ha/bohr, below the previous 5e-4 floor. A separate diagnostic passes its unchanged 1e-9 finite-difference accuracy requirement: the corrected error is 6.7e-13 Ha/bohr, while omitting grid motion gives 1.27477e-4 Ha/bohr. The replacement non-vacuity condition requires the omitted-correction error to exceed 1,000 times the accuracy tolerance, rather than preserving the old partition’s correction magnitude. Its first attempt hit a 16-GiB memory cap at 17.634 GB; the unchanged diagnostic completed under a 24-GiB cap with an 18.411-GB observed peak. grid-motion-witness-review.json retains the force tensors and provenance. The gate’s source files stayed unchanged until the full run and separate MgO test completed. Only the two reviewed test functions changed for their reruns; the rest of both modules was compared structurally and is unchanged. The selection retains two baseline markers: a strict expected NotImplementedError for padded multi-k RHF CPHF negative curvature, and a runtime-budget skip for the MgO corrected-Gamma RHF gradient. The expected failure exercises historical padded domains with pair_complete_1e=False; physical coupled SCF finite-difference cases are maintained separately. Their decorators were compared structurally with standalone base bdec7d2 and are unchanged. They must be reported separately from passing tests. The MgO row was also run independently with eight native threads, removing only its runtime skip while retaining its assertions and settings. Its monitor was also extended under the same 16-GiB limit, retaining the exact kernel-reported exit status. verified/budgeted-guard-handoff.json distinguishes the monitor replacement from the pytest outcome. At 03:16 UTC on September 9, the allowance was extended to six hours total, ending at 05:55:54 UTC, because the fixture’s recorded historical runtime was 5.4 hours. The same pytest process continued under the same memory cap. The replacement armed its kernel exit watch before stopping the old monitor; that monitor’s exit 137 is distinct from the eventual test exit. The first extension receipt remains in budgeted-guard-handoff-first-extension.json.

The CRYSTAL CYC0 component fixture now injects its sealed local SAD density explicitly. Production guesses normalize in the periodic overlap metric, whereas the sealed diagnostic input used the molecular metric. Fixing the diagnostic input preserves the independent CRYSTAL values and the checks that distinguish CYC0 from the returned terminal density. The renamed fixed-density component test passes; production guess normalization is unchanged.

For XC convergence, an independent contraction of finite Bloch AO image sums matches the native physical-domain energy to 2.9e-12 Ha. Both bra and ket image domains must be specified: the default native bra reach is 10 bohr and the ket reach here is 20 bohr. The unrestricted periodic AO sum is a different diagnostic. Native and PySCF XC functionals agree on the same density and gradient. The energy-only contraction avoids forming the XC potential matrices and is used for the subsequent quadrature ladder.

At partition radius 18 bohr and fixed 10-bohr bra support, increasing quadrature from 75/17/36 to 150/25/50 and 200/35/72 gives XC energies -2.262343324060281, -2.262383493413342 and -2.262382287809644 Ha. On the finest grid, expanding bra support to 14 and 20 bohr gives -2.262380998300450 and -2.262380998501048 Ha. The final value differs from the independent level-9 XC reference by +0.87709 microhartree; the last support increment changes it by only -2.01e-10 Ha. This completes the energy-only XC ladder, not a native SCF convergence ladder.

The fresh Si/PBE reference is converged at grid levels 5, 7 and 9: -578.6580706412801, -578.6580370246670 and -578.6580312081404 Ha. The last grid increment is +5.81653 microhartree. All levels use the same 44-function basis, 28 electrons and zone-centered 2x2x2 mesh. Level 9 reuses the analytic one-electron matrices from levels 5/7 and rebuilds the independent RSJK operator. This completes the previously interrupted reference calculation; the native Si comparison is recorded below.

At the level-7 Si reference density, changing only the native grid partition from the historical image weights to the corrected point weights shifts XC by +20.52974479 mHa. Both evaluations use 75 radial, 17 polar and 36 azimuthal points per atom and a 10-bohr partition radius. The integrated electron count changes from 28.08477868 to 27.99970363. Increasing the corrected partition radius to 18 bohr changes XC by -48.71248 microhartree; increasing quadrature there to 150/25/50 changes it by another -2.37605 microhartree. These are fixed-density component comparisons, not converged native SCF results. An independent check of all 44 Si AOs at 800 sample points agrees with the native basis to 7.2e-15 absolute error after the documented angular-order permutation.

At the fresh level-9 Si density with matched 18-bohr partition and bra support, quadratures 150/25/50 and 200/35/72 yield XC energies -41.6639507896482 and -41.66395319182573 Ha. The last increment is -2.40218 microhartree and the finest value is +5.68555 microhartree from the independent reference. Its integrated electron count is 27.99999915. The ket-image support remains 20 bohr. These fixed-density results retain the independent-basis and native-functional cross-checks.

Increasing physical pair and nuclear cutoffs from 20 to 24 bohr at the same LiH reference density changes the non-XC energy sum by +1.47e-11 Ha. The automatically resolved SR interaction reach increases from 35.47199 to 39.47199 bohr. At 24 bohr, all non-XC components agree with the independent reference within 3.7e-12 Ha, and J changes from its corrected 20-bohr value by +5.34e-12 Ha. The 20-bohr comparison combines the recorded one-electron witness with the separately corrected J witness; the archive identifies both inputs explicitly. This is a fixed-density cutoff check, not another SCF ladder. Each driver normalizes the supplied reference density in its finite overlap metric; the receipts record those factors, which approach unity as support increases. For both LiH and Si, expanding only ket-image support from 20 to 24 bohr on the default grid leaves XC energy and integrated electron count bit-for-bit unchanged.

The corresponding Si 20-to-24-bohr check changes the non-XC energy sum by -2.92e-10 Ha. Its largest individual component change is 2.45e-9 Ha; all 24-bohr non-XC components agree with the independent reference within 2.13e-9 Ha. J is 174.42057194564217 Ha and the resolved SR interaction reach is 38.98511 bohr. These runs used four native workers under concurrent host load; their 587.75-second LiH and 1549.29-second Si timings are not controlled performance measurements.

The pre-split 175 passes plus one failure and separate 17 passes were interrupted sessions. They are historical partial receipts, not completed gates for this checkout. The development environment is rebuilt locally; new run inputs and logs are held under /tmp/bipole-erfc-resume/ while the continuation is active. Durable snapshots, including full patches and checksummed inventories, are preserved separately under ~/vqruns/bipole-erfc-20260908-standalone/. The original handoff archive is unchanged.

The native LiH/PBE SCF, warm-started from the independent level-9 density, converged in four iterations and passed a cold exact-Fock confirmation. Its total energy is -8.103446644267258 Ha, +33.99929 microhartree from the independent result. The exact commutator norm is 1.356e-7 and the confirmation changes the loop energy by -4.893e-11 Ha. This is consistent with the +34.01692-microhartree default-grid XC discrepancy measured at fixed density. The run used the default 75/17/36 quadrature and 10-bohr partition radius, a 20-bohr pair/nuclear cutoff, and the resolved 35.47-bohr SR interaction reach. It took 2197.55 seconds with one native worker; the reference density is a warm start, so this is not a cold-start or performance benchmark.

LiH loaded the Schwarz-corrected extension de0e5158... before the screened-derivative fix was installed atomically. The already-running process retained that loaded image; its SCF did not call the changed gradient routine. The raw result’s hash was sampled from disk at return and therefore names the newer file. lih-native-scf-verified.json corrects the loaded-image field while preserving that raw hash and linking the replacement receipt. The raw result is retained unchanged.

At the level-9 Si reference density, native kinetic, nuclear-attraction, nuclear-repulsion and Coulomb components all agree with the independent values within 4.0e-9 Ha. Native J is 174.42057194496707 Ha, a difference of -2.02e-9 Ha. This check used the final native build and four workers; it took 254.00 seconds.

The full native Si/PBE SCF also converged in four iterations and passed its cold exact-Fock confirmation. Its energy is -578.6579424366462 Ha, +88.77149 microhartree from the independent level-9 result. The exact commutator norm is 3.682e-7; the confirmation changes the loop energy by +2.956e-12 Ha. The final density evaluation reuses this confirmed build. The run used the default grid, a 20-bohr pair/nuclear cutoff, a resolved 34.9851-bohr SR interaction reach, and eight native workers. Its loaded extension hash was captured at startup and matches the final build. Wall time was 783.20 seconds, with a peak observed RSS of 5.61 GB. As for LiH, the independent density is a warm start and this is not a timing benchmark. At the exact level-9 reference density, the default-grid Si XC value is -41.66387009198551 Ha, +88.78539 microhartree from the fine reference. Together with the non-XC component check, this accounts for the converged energy difference; the remaining density relaxation contributes about -0.0122 microhartree. The default-grid integrated electron count is 27.99970363. The finer-grid/support ladder above reduces the fixed-density XC discrepancy to +5.68555 microhartree.

Several early diagnostic scripts accidentally shared one result filename. Their scalar logs survive, and the final Si and corrected LiH matrix snapshots were recovered into distinct files. Earlier full matrices could not be recovered and are not claimed as available. The archive’s native-result-recovery.json identifies the affected runs, authoritative outputs and retained scalar evidence. Exact executed scripts are kept separately from corrected reproduction scripts.

Outstanding acceptance work includes broader crystal convergence checks and a controlled performance review. The padded multi-k RHF CPHF limitation is resolved by the follow-up below. The LiH and Si 20-to-24-bohr fixed-density cutoff checks above are complete. The global physical-domain default stays off pending that broader review. Terminal XC reuse and the replacement standalone gate regression have completed their local validation.

The complete 840-case run reconciled on September 9 in 9,965.81 seconds, with a 20.895-GB peak observed RSS. All setup/call/teardown and JUnit identities agree, with no omitted or duplicate cases. All 568 native inputs, 600 Python sources and 17 selected test files matched their recorded hashes. The two failures are exactly the grid assertions reviewed above; the raw full-run result remains a failed run. See completed-gate-reconciliation.json.

The reviewed-fixture rerun passed both cases under the same native binary. The legacy RKS pin now uses a strict absolute 1e-9-Ha tolerance (relative slack disabled) and explicitly checks its unchanged spheropole. The moving-grid force accuracy remains 1e-9 Ha/bohr; omitting the correction must fail by at least 1,000 times that tolerance. The complete run plus these two reruns establish 838 passing selected cases, with the one known padded CPHF expected failure and one unchanged historical runtime skip reported separately. final-gate-reconciliation.json preserves the separate full-run and rerun counts.

The historically skipped MgO corrected-Gamma RHF gradient passed all test phases independently in 19,163.01 seconds with eight native threads. All three SCFs converged, net force satisfied its bound, and the nonzero oxygen force matched the central finite difference without changing settings or tolerances. The exact process exit was zero; see budgeted-gradient-reconciliation.json.

September 9 follow-up: padded RHF response after 294d3e5

The former strict expected failure for padded multi-k RHF now passes as an ordinary regression. Its asymmetric BeH2/STO-3G [2,1,1] fixture retains the fold-reliable 13-bohr cutoff, all SCF settings, and the complete analytic-vs-FD assertion body, including the 5e-5-Ha/bohr tolerance. No native code changed.

The previous finite-difference CG callback failed at iteration 42 with p^T A p = -1.715e-21. Explicit unit-direction assembly gives a full-rank 48-dimensional response matrix; CG on that stored matrix converges in eight iterations. In consistent complex orbital coordinates, its symmetric-part minimum eigenvalue is 0.2600601 and its condition number is 20.69282. These observations support a numerical callback failure rather than the former exception’s implication of an SCF instability.

The corrected response assembles the Jacobian using unit rotations, solves its transpose, and includes the k-point weights in the energy RHS. Occupied rotations use the conjugate-transpose convention matching the complex occupied-virtual residual. The final contraction does not apply those weights again. The assembled equation must satisfy the existing 1e-9 relative residual tolerance; inconsistent singular systems still fail closed. Dense matrix assembly remains a scaling cost, so this does not certify large-system gradient performance or change the preview status.

Five small-model cases differentiate independently constructed normalized AO density energies along a coupled SCF constraint. The four real/complex and uniform/unequal-weight cases fail before the change and pass after it; the singular inconsistency remains refused. Test-only forward and symmetrized solves fail all four derivative cases. Restoring the old complex convention or applying weights after solving each fails two cases.

The patched production process passed all seven selected cases: those five models and both padded/unpadded BeH2 native gradient regressions. The separate fast gradient and test-gate run passed 93 cases, giving 100 unique targeted passes. Per-phase events and JUnit identities agree; the production source hashes stayed unchanged throughout the run. A separate old-code diagnostic with an explicit solver override passed its BeH2 finite-difference check at 7.07e-8 Ha/bohr maximum error; it is not counted as a production pass.

The local continuation archive bipole-erfc-20260909-cphf contains acceptance-summary.json, operator-analysis.json, the raw native and model results, the original expected-failure receipt, and a full patch against 294d3e5. Its inventory and isolated patch reconstruction preserve exact source provenance. The earlier 840-case record remains unchanged; this follow-up does not relabel that historical full run as green. Broader crystal and cold-start/grid SCF checks, controlled performance/default review, and fleet/tracker closure remain outside this targeted follow-up.

September 9 crystal convergence and default review after 70aa93a

This follow-up checks LiH rocksalt and Si diamond with the same PBE, pob-tzvp-rev2 basis and unreduced [2,2,2] mesh as the earlier independent level-9 comparisons. The native runs start from HCORE, with no supplied reference density. They use physical 20-bohr pair/nuclear support, omega 0.4, Ewald precision 1e-12, Schwarz threshold 1e-14 and SR image precision 1e-8. The grid retains the default 75 radial by 17 theta by 36 phi points and 10-bohr image radius. The energy and commutator exit criteria remain 1e-9 Ha and 1e-6; the iteration cap is 80. Two native threads run the crystals sequentially on a shared host, so their wall times are validation costs rather than a performance benchmark.

Both cold starts passed the exact, non-incremental terminal Fock check and reused that confirmed build for final density evaluation. Their energies agree with the earlier reference-density warm starts well inside the predeclared 1e-8-Ha agreement tolerance. All production Python source files and the native binary matched their recorded startup hashes at return.

Crystal	Iterations	Energy (Ha/cell)	Cold minus warm (Ha)	Exact commutator
LiH	7	-8.1034466442673	-4.441e-14	5.581e-09
Si	10	-578.6579424366507	-4.547e-12	4.988e-08

The maximum AO density-element differences from the warm starts are 8.784e-09 for LiH and 1.566e-07 for Si. The occupied subspaces are reconstructed from the saved native MO coefficients, with positive indirect gaps checked before assuming equal integer occupations at every k point. Wall times are 4178.05 and 4443.24 seconds; the combined guard’s peak observed RSS is 6.589 GB.

The separate generic fixed-density grid survey uses the independent level-9 PySCF density and finite-image AO sums on the native grid points and weights. Bra-image reach equals the partition radius; ket-image reach stays at 20 bohr. It does not wrap points or substitute unrestricted periodic AO sums, which would change this finite operator. Native and PySCF LibXC evaluations of the resulting density agree within 1e-9 Ha in all 16 cases. Both default-grid values also reproduce the earlier sealed calculations within 5e-10 Ha. All recorded input hashes match.

Grid	Radius (bohr)	Points/cell	LiH XC error (µHa)	Si XC error (µHa)
default	10	91,800	+34.017	+88.785
leb100-29-r14	14	60,400	-216.729	-437.482
leb150-35-r14	14	130,200	-72.364	-185.281
leb200-41-r14	14	236,000	-24.562	-72.011
leb200-41-r18	18	236,000	-25.953	-84.954
leb200-47-r14	14	308,000	-9.189	-34.460
leb200-53-r14	14	389,600	-2.006	-12.305
leb200-53-r18	18	389,600	-2.621	-16.003

Here lebN-L-rR means N radial points, Lebedev order L and image radius R bohr, with no angular pruning. Errors are relative to each independent level-9 XC value at its fixed density, per primitive cell. These are quadrature comparisons, not self-consistent fine-grid energies.

The archived LiH reference used PySCF 2.6.2; Si used 2.14.0. Their “level 9” labels therefore do not specify identical radial and density-pruning defaults. The explicit historical LiH radial formula and pruning rule reproduce its 189,282 points and XC value within 3e-15 Ha under the current runtime. Current 2.14 settings give 313,990 points and shift fixed-density LiH XC by -0.062443 microhartree. This explains the version-sensitive reference check; it does not certify either quadrature as converged.

The small Lebedev rules do not justify replacing the product grid: order 29 is worse for both crystals despite using fewer points. At 200 radial points and radius 14, orders 41, 47 and 53 progressively reduce the discrepancy. Increasing radius 14 to 18 at order 53 changes XC by -0.616 microhartree for LiH and -3.698 microhartree for Si. This identifies an improved grid candidate, but does not certify the grid limit or replace the earlier, finer product-grid evidence. A complete native fine-grid SCF ladder remains outstanding.

Four additional cases used the bundled PySCFLevel3 atomic-grid profile, which overrides the generic radial/angular/pruning controls as a complete protocol. It produces 23,530 points for LiH and 38,360 for Si. At radii 10/14 bohr the LiH XC errors are -241.183/-299.164 microhartree, and Si’s are -2.385/-182.610 microhartree. The apparent Si agreement at the smaller radius is not stable under increased support: its integrated electron errors are +0.00058179/+0.00139717, both larger than the default product-grid error magnitude of 0.00029637. This cheaper profile therefore does not justify a default replacement. These four cases also passed native/PySCF LibXC agreement and source-hash checks, bringing the complete fixed-density survey to 20 cases.

The actual defaults were read from the imported native options and Python driver signature. The validation overrides are material:

Setting	Direct RKS API default	Cold-start validation
Physical pair domain	Off	On
Pair / nuclear cutoff	15 / 25 bohr	20 / 20 bohr
Ewald omega	Automatic	0.4
Ewald precision	1e-8	1e-12
Schwarz threshold	1e-12	1e-14
SR image precision	1e-6	1e-8
Initial guess	AUTO	HCORE
Energy / commutator tolerance	1e-8 / 1e-6	1e-9 / 1e-6

The runner can resolve cutoffs differently; this table describes the direct API used here. In physical mode at a fixed 20-bohr pair cutoff and omega 0.4, the default SR precision resolves interaction radii of 33.2066 bohr for LiH and 32.7910 bohr for Si, versus 35.4720 and 34.9851 bohr at the tighter validation precision. The radius policy also enables charge-pair Schwarz screening. Its precision parameter is a radial-tail estimate, not a bound on the total lattice error.

The fixed-density native comparison changed all three screening precisions back to their defaults, holding the physical 20-bohr pair/nuclear support and omega 0.4 fixed. The non-XC energy changes relative to the tight settings are -1.55e-12 Ha for LiH and -9.61e-10 Ha for Si. Maximum Coulomb-matrix element changes are 1.90e-8 and 2.89e-9 Ha, respectively. Every measured default-screening non-XC component remains within 3.96e-9 Ha of its independent reference. These changes are much smaller than the default-grid XC discrepancies. They do not certify automatic omega, the default cutoffs or a fully default SCF.

The first screening recorder failed after the LiH native build because it read driver settings from the nested Fock callback’s keyword arguments. That attempt supplied no accepted J receipt; its owned Si child was stopped when the shared recorder defect was identified. The corrected callbacks passed isolated checks before both numerical probes were rerun. Raw failed and completed attempts are preserved separately.

The read-only fleet inventories captured for this follow-up do not attest this branch: compute-cluster’s registered development runtime reports source 5ad944fbeffe and version 0.15.162; compute-managed’s executable registrations expose no source SHA. No cluster calculation or deployment was used for these results. The issue snapshot records all five migrated issues as open: #187 (old #20), #186 (old #21), #182 (old #66), #82 (old #674), and #44 (old #724). The two-crystal checks do not cover the complete affected crystal set, automatic-omega and cutoff defaults, or controlled HF/hybrid performance. The global physical-domain switch remains off; no tracker closure is claimed.

The bipole-erfc-20260909-crystal-convergence archive records the exact executed scripts, startup provenance, raw logs, grid tables, screening comparison, cold/warm reconciliation and required earlier references. Its full patch reconstructs the committed tree from the standalone base, and its inventory verifies every archived file.

Rebased native fine-grid follow-up (2026-09-09)

This follow-up rebased the continuation onto standalone main ed5deffae394ec195fdb2f000b4779728e85a1f3 and rebuilt the native extension. Its SHA256 is 089312df98c2a410b75b46f9a569a487039b2e81a0936940f8f0f4dff1d672ba. The rebased integration checks passed 181 binding/domain/terminal/CPHF tests, 506 upstream physical-product/finite-source/Gram tests, and the two maintained padded/unpadded multi-k RHF finite-difference force gates. These 689 checks are separate from the historical raw 840-case run described above.

The rebuilt default-grid LiH and Si continuations converge in two iterations and reproduce the prior HCORE solutions within 8e-12 Ha. Independent PySCF/LibXC XC contractions agree within 8e-12 and 4.5e-9 Ha respectively. Fresh independent RSJK Coulomb operators plus the sealed one-electron matrices reproduce total energies within 1.3e-11 and 2.3e-9 Ha, and Fock matrices within 1.4e-9 and 3.4e-9. These default checks reconstruct the returned canonical-MO densities; they do not claim to use the exact accepted lattice density, which the initial default recorder did not serialize.

The first fine-grid attempt exposed an AUTO-domain radius propagation bug: the quadrature used the requested 14-bohr radius but XC’s AO bra images still used the unset native 10-bohr default. Commit 6dee928 propagates the requested periodic-grid radius in RKS and UKS. At the saved LiH seed, independent XC reproduces the defective iteration’s energy within 5e-11 Ha, and increasing only the bra radius from 10 to 14 shifts XC by +1.3043 microhartree. The partially run defective case is retained and excluded from fine-grid acceptance. The fix passed 144 terminal/provider tests and 47 KS integration tests.

The corrected native fine-grid protocol uses 200 radial points, unpruned Lebedev order 53 and radius 14 (389,600 points per cell), physical pair and nuclear cutoffs of 20 bohr, omega 0.4, Schwarz 1e-14, Ewald precision 1e-12 and SR precision 1e-8. Every intercepted native XC call records and checks its actual 14-bohr bra radius and 20-bohr density cutoff. The recorder saves the accepted lattice density, its validated per-k representation, and the final Fock/overlap/hcore matrices. Canonical orbitals are retained separately.

Native LiH converged in four iterations to -8.103482649513941 Ha/cell. Its exact accepted-density commutator is 1.464e-7; the exact rebuild differs from the loop energy by +4.831e-11 Ha and is reused for final evaluation. All five XC calls use the requested 14/20-bohr support. Independent XC at the saved accepted density agrees within 2.3e-15 Ha, and the independent full operator agrees within 1.1e-11 Ha in energy and 8.3e-10 in the Fock matrix. The density round-trip error is below 6e-17.

A separate PySCF AO/LibXC/RSJK SCF loop, with its own DIIS, diagonalization and stopping rules, starts from the independent level-9 density. It shares only the native quadrature and uses sealed reference one-electron matrices. For LiH it converges in four iterations to -8.103482649503272 Ha/cell, with residual 5.797e-8. Its energy differs from the native result by 1.1e-11 Ha; the orthonormal-density difference is 5.9e-8. The corrected fine-grid native LiH result is 36.005 microhartree below the default grid and 2.006 microhartree below the finite level-9 reference.

Native Si converged in five iterations to -578.6580435165043 Ha/cell. The exact accepted-density residual is 1.199e-8, and the confirmation changes energy by +5.093e-11 Ha. Five loop XC builds plus one confirmation are recorded, with no additional terminal build. The density round-trip error is below 5e-16. Independent accepted-density XC agrees within 3.6e-14 Ha; the independent full operator agrees within 2.3e-9 Ha in energy and 2.1e-9 in the Fock matrix.

The independent Si loop converges in six iterations to -578.6580435142816 Ha/cell with residual 1.327e-8. Native and independent energies differ by 2.3e-9 Ha, and the orthonormal-density difference is 1.6e-8. The fine-grid native result is 101.080 microhartree below the default grid and 12.308 microhartree below the finite level-9 reference. Thus both independent loops reach the same fine-grid solution within the predeclared energy/density/residual gates. They do not establish the infinite grid limit. Native driver wall times are 5844.05 seconds for LiH and 6851.46 for Si, under shared host load; these are not controlled benchmarks.

The native Si case was started concurrently with LiH using a byte-identical script and the same four-thread, four-hour and 24-GiB case limits. The original serial controller’s later duplicate was stopped and its -15 exit preserved. Accepted-case assembly requires a zero exit from the concurrent Si controller and explicitly excludes that cancelled duplicate. The raw serial campaign is retained as incomplete; it is not relabelled successful.

The associated gradient radius correction is committed as c0c3737, with test-only strengthening through 529c2a1. It copies lattice options without changing the caller’s fields and propagates the grid radius through the fixed-grid and moving-grid XC terms. The new tests fail in eight cases on the old source and pass all 12 with the correction. The preservation regression enumerates all 13 public native option fields independently of the helper. A deliberate omitted-field mutant fails in both periodic KS cases, while all eight unmodified preservation combinations pass.

The checkout’s own native environment completed the four maintained Gamma LDA/PBE RKS/UKS force witnesses and the remaining slow matrix, with 42 passes and one historical skip. These processes loaded the exact committed gradient module without changing the package files held by concurrent native SCFs. All startup source hashes match at return. Every preexisting test function and fixture is AST-identical to the committed final test file.

After integration, the checkout’s own native environment passed all 57 fast gradient cases in 147.28 seconds. Combining those with the four Gamma KS witnesses, 42 remaining slow passes and two rebased padded/unpadded RHF witnesses accounts for every current gradient test exactly once: 105 pass and one historical high-cost MgO skip, across 106 collected nodes. No preexisting test tolerance was relaxed. The skip is not promoted to a pass on this binary; its prior unskipped result belongs to the earlier archive. The complete follow-up contains 965 distinct passing test nodes across the rebased integration and radius/gradient gates, with that one skip. This is selected-gate evidence across the recorded commits, not a full-inventory run on a single final SHA.

H2 RKS LDA/PBE force errors after that correction are below 3e-9 Ha/bohr; H3 doublet UKS errors are about 5e-6, versus errors above 1e-3 before the fix. The exploratory 1e-7 criterion failed for UKS and remains a recorded failure. The smaller remaining error depends on the preview’s internal grid-motion finite-difference step. The permanent UKS regression uses 1e-5, checks two smaller relaxed-SCF difference steps, and changes no existing tolerance. This does not broaden analytic-gradient certification.

A separate fixed-state probe enables physical domains but otherwise uses direct-API cutoff/screening/automatic-omega defaults. The shared options clamp the requested nuclear cutoff from 25 to 15 bohr; private Ewald source support is resolved separately. The non-XC sums differ from the independent reference by -0.224 microhartree for LiH and -149.872 for Si. The Si input electron trace is 27.99997991996 before normalization; its scale factor is 1.000000717145. These component differences include that density normalization and cannot be attributed to a single cutoff term. This one-build probe stops before XC and is not a full default SCF. It reinforces the decision to retain opt-in physical domains and to avoid promoting the current cutoff/grid defaults.

An off-tree finite-Bloch XC factorization reproduces native default-grid energies within 3e-14 Ha and raw potential matrices within 2.5e-13 on both crystals. Non-TRIM complex-density controls discriminate the phase convention; directional derivative and batch-size checks also pass. One-thread observed wall times are 72/345 seconds for LiH and 76/315 for Si (prototype/native), under shared host load. This is neither a controlled benchmark nor a production optimization. A production path still needs validated accepted per-k densities, bounded AO batching, and separate spin/meta-GGA/external provider/gradient coverage.

The final bipole-erfc-20260909-native-fine-grid archive contains the four accepted native SCFs, authoritative fine-grid densities, independent XC/operator/loop checks, gradient collection and coverage accounting, expected negative controls, source snapshots, and the exact tested binary. Its continuation patch reconstructs the committed tree from the standalone base, and its inventory hashes every archived file. Local acceptance does not close the global physical-default, complete crystal-set, asymptotic-grid, HF/hybrid performance or independent-person verification gates. No cluster run, deployment, tracker mutation or upstream push was used.

Explicit reference and angular-grid refinement, 2026-09-09

The follow-up separates radial, angular and image-radius effects at both the saved native accepted density and the archived independent reference density. All nine profiles per material are accounted for, giving 36 primary density evaluations plus the repeated controls. Failed attempts remain separate from completed continuations and their repeated controls. Both time-guarded attempts are resumed with an equivalent optimized contraction order, with real/complex controls and a repeated native-grid check. Native and PySCF LibXC energies agree within 1e-9 Ha throughout. The initial LiH current-version reference check remains a recorded failed attempt; its historical-grid reproduction and the version-aware rerun are separate receipts.

Five additional independent PBE reference SCFs use explicit atom-specific Treutler radial grids, NWChem angular pruning and no density pruning under PySCF 2.14.0. The unreduced [2,2,2] meshes, bases, sealed analytic one-electron matrices and fresh exact RSJK Coulomb builder remain fixed. Every accepted state passes the fresh final Fock check, with maximum commutator 5.689e-10. Source hashes remain unchanged.

Radial / angular points	LiH energy (Ha/cell)	Si energy (Ha/cell)
300 / 2030	-8.103479811961442	-578.6580270441226
400 / 2702	-8.103479770512775	-578.6580259849007
500 / 3470	Not run	-578.6580257401564

The last reference refinement changes LiH by +0.041449 microhartree and Si by +0.244744. Si’s extra 500/3470 calculation follows the 1.059222-microhartree change between the first two refinements. These are measured steps, not an exact-limit bound.

Two independent finite-image SCFs replace the 200/L53 angular rule with a 200-radial, 35-theta, 72-phi product grid at the same 14-bohr bra/partition radius and 20-bohr ket support. They start from the previously verified independent Lebedev-grid states. Their AO evaluation, XC contraction, RSJK, DIIS, diagonalization and convergence logic are independent of the native BIPOLE SCF; the native quadrature and sealed reference one-electron matrices are shared. Both pass the 1e-10-Ha energy-change and 1e-7-commutator criteria.

Crystal	Independent product-grid energy (Ha/cell)	Minus last reference (µHa)
LiH	-8.103479620614127	+0.149899
Si	-578.658026129749146	-0.389593

The first product-loop energies match the fixed-density grid-change predictions within 1e-8 Ha. Subsequent density relaxation changes LiH and Si by -0.000115 and -0.001396 microhartree, respectively. The following controls use the saved native accepted densities, with all other parameters held fixed unless stated.

Grid change	LiH XC change (µHa)	Si XC change (µHa)
Lebedev 53: radial 200 to 300, R14	-0.527701	-3.072054
Lebedev 53: radius 14 to 18, radial 200	-0.615610	-3.698320
L53 to product 35/72, radial 200, R14	+3.029003	+17.385929
Product 35/72: radius 14 to 18, radial 200	-0.145861	+0.607076
Product 200/35/72 to 300/47/96, R18	+0.135633	+0.096939

All numerical workers finish before the record is sealed. The bipole-erfc-20260909-grid-limit archive preserves scripts, failed and completed attempts, authoritative densities, current and historical reference sources, runtime hashes and the continuation patch. The production extension remains the previously tested 089312df binary.

The completed native SCFs still use 200/L53/R14; the new product-grid SCFs are independent loops, not additional native SCF acceptances. This study does not promote grid/cutoff defaults or close broader crystal, HF/hybrid performance, independent-person review or global physical-domain gates.

Midpoint-aware nuclear support, 2026-09-09

The physical nuclear selector centres its source ball on the arithmetic AO-pair midpoint. The Python Ewald helper previously retained legacy cell-origin padding, so merely relabelling one atom by a lattice vector unnecessarily enlarged that ball. For positive primitive exponents, the Gaussian product centre lies at most half the pair separation from the midpoint. The physical branch now adds half the pair cutoff to the smeared erfc range. The legacy branch and explicit larger source cutoffs are preserved. Energy and analytic derivatives continue to use the same helper.

Three atom-image relabelling controls fail on the old helper and pass with the correction. All 24 nuclear tests pass, including 1D/2D/3D labels, diffuse and asymmetric s/p/d products, larger-domain energy/gradient comparisons and caller-option preservation. No native extension rebuild is required: the tracked C++ sources and tested 089312df extension are unchanged.

The saved native fine-grid LiH and Si densities also compare the exact old helper with the corrected helper, using the same native extension. These are fixed-density nuclear-component witnesses, not new native SCFs.

Crystal	Old / midpoint radius (bohr)	Nuclear energy change (Ha)	Max matrix change	Max gradient change
lih	42.954732 / 26.271058	-1.776e-15	1.488e-14	4.770e-17
si	40.358462 / 25.914408	+2.274e-13	4.707e-14	7.262e-16

The two H2/SVWN retirement pins are corrected separately. Both failed on the unchanged checkout. Restoring only the historical grid partition reproduces both old pins within their original 5e-9-Ha tolerance; the corrected partition shifts both routes by +4.05943148 microhartree entirely in XC. The two numeric anchors are the only AST changes in that test file. All eight tests pass on rerun, with unchanged bodies and tolerances.

The 67-file fast BIPOLE lane initially records 1,906 passes and three failures. The reviewed retirement rerun resolves two of those failures. Two files have only slow cases and therefore no selected fast tests; 31 non-gradient slow cases remain deselected across the lane. The explicit slow-gradient batches cover all 49 additional gradient nodes. Together with the 57 fast nodes, the gradient matrix has 105 passes and one historical expensive MgO skip. The original 14-GiB-guarded batch returned no accepted tests; its complete 18-node rerun is recorded separately under a 24-GiB/four-hour guard.

The other two original gradient workers retain their processes, threads, source and 14-GiB limits. Their local time budget is extended from two to four hours, measured from their original starts, through an overlapping supervisor handoff. Darwin exit-status observation is checked against normal, nonzero and signal exits before use. The original supervisor is recorded as superseded; its old time budget is not retroactively labelled a successful bounded run.

After exact accounting, these selected checks contain 1,980 distinct passes, one historical gradient skip and one unresolved existing SAP failure. This is not a clean full repository gate or independent-person verification. The raw lane failures and stopped resource-guard attempts remain in the archive.

test_vsap_gamma_symmetric_and_attractive fails identically on the unchanged checkout and candidate: raw SAP Gamma asymmetry is 0.001311559 against 0.001. The midpoint helper is not called. Restoring the historical partition gives 0.000108793; current radius-14/radius-18 and physical-domain controls still exceed the unchanged threshold. No test tolerance, skip or production default is changed to absorb this failure. SAP/grid analysis remains open alongside broader physical-default, crystal-set, HF/hybrid performance and other-person verification gates.

SAP quadrature and upstream integration, 2026-09-12

The continuation is rebased onto standalone core main 3a5f1d5. The merge resolution retains the new ECP SAP dispatch inside physical ERI-cell mapping, and keeps both the upstream MDF and BIPOLE test-lane assertions. The native extension is rebuilt against this checkout’s vendored dependencies; its SHA256 begins 263b78c8. The September 9 source and numerical archives remain separate. Their 1,980-pass count and refined SCF values are not re-certified for this rebased tree.

The previously unresolved SAP symmetry test exposed a quadrature contract mismatch. Each SAP lattice matrix block integrates one localized AO pair over all space, while the shared Python caller supplied periodic per-cell weights without summing bra images. The smooth potential is periodic; the localized AO product is not. The caller now uses the same molecular Becke partition as the native SAP drivers. The direct integral tests use that contract as well. Neither the periodic potential nor the existing symmetry threshold is changed.

Two new regressions inspect the grid at the real Python SAP call boundary. For a constant potential, its quadrature must reproduce the analytical localized overlap within 2e-6. They also check the raw SAP matrix before the helper’s existing Hermitian projection. Fresh-core diagnostics give:

LiH cell edge (bohr)	Periodic-weight overlap error	All-space overlap error	SAP asymmetry, periodic / all-space
7.6	0.289839	7.951e-8	0.0346592 / 0.000229442
12.0	0.0712418	4.226e-7	0.00131156 / 0.000108793

The rebuilt baseline fails both new regressions and the old symmetry test. A stale generated basis overlay was then refreshed. Its STO-3G file bytes changed, so all three failures were reproduced again with the refreshed basis data before testing the fix. The SAP table stayed byte-identical. Final test receipts attest the refreshed overlay alongside Python source and the rebuilt extension.

The completed targeted groups record 285 distinct passes. They include the full SAP test file, explicit open-shell SAP caller checks, 150 physical/nuclear cases and 93 binding, lane and caller-integration checks. A broader 189-case SAP/open-shell attempt is stopped during GDF Fourier-factor setup for a NaH ECP AUTO/SAD case, before reaching its guess builder. Its partial passes are not counted; the relevant SAP cases are rerun separately. That broader attempt also exposes a PATOM first-cycle comparison failure. The PATOM failure reproduces with the pre-fix helper restored in memory and all SAP construction calls forbidden; that route never enters the changed construction. Its first-cycle energies are -1.198539021862297 Ha (GDF) and -1.196328119109694 Ha (direct Ewald), a -2.210903-mHa difference. These are one-cycle states, not converged SCFs. The comparison remains unresolved, with the assertion unchanged.

These are targeted integration results, not a full BIPOLE lane or complete gradient-matrix rerun. The historical MgO gradient skip, broader physical-default and crystal-set work, and independent-person review remain open. The separate bipole-erfc-20260912-sap archive preserves the rebase, failed baselines, completed results, source and basis hashes, rebuilt binary and exact-tree continuation patch. No upstream push, fleet operation, tracker mutation or issue closure is performed.

Standalone issue review, 2026-09-13

The continuation was rebased onto standalone main 9010c00, producing e654cb8. All 576 recorded native build inputs match the September 12 vendored build, extension SHA256 263b78c8f53a13c2a47d78c3e1aa2ef7dc4d3d54b50b7ca9191687f1c57e6289. The rebase refreshed source timestamps, so the timestamp-only stale-core warning is present; the native content comparison has no mismatches.

Standalone issue #216 groups the retirement and legacy-gauge reference shifts. Six new converged SCFs repeat the paired partition experiment on this native build. Only the grid partition changes within each pair.

Route	Historical partition (Ha)	Point-centered partition (Ha)	Change (microhartree)
SR/LR	-1.1211961982135232	-1.1211921387820452	+4.059431478
Exact FT	-1.1211968598808090	-1.1211928004493320	+4.059431477
Legacy gauge	-1.1156375080452340	-1.1156334486137565	+4.059431478

Every historical value recovers its original pin within the original absolute tolerance. Kinetic, nuclear-attraction, nuclear-repulsion and J/K components are unchanged to 1.2e-16 Ha; the energy change is wholly XC. The exact-FT minus SR/LR gap stays -0.661667286 microhartree within 9e-16 Ha. The legacy spheropole is exactly 0.005559511758728281 Ha in both runs. These controls establish the shared partition cause without a gauge offset or altered acceptance tolerance. The two affected test files pass on this build.

Additional completed checks give 125 passes for native binding, gate, Schwarz and nuclear coverage, and two passes for the H2/H-atom BIPOLE-versus-Ewald controls at the existing 1e-7-Ha tolerance. The broad fast and slow runs are distinct from these completed selections. The fast run encounters a per-test timeout in the wide reference build of test_padded_internal_cells_reproduce_wide_build; its raw failed receipt has no completed pytest count. A shared resource hold prevents queued files from starting. Neither held targets nor partial output count as passes, and the old four-file known-red entry is not retired on this basis.

Issue #220 now tracks the September 12 all-space SAP quadrature correction. The separate PATOM first-cycle diagnostic remains unresolved: suppressing only GDF’s one-electron cutoff growth from 7 to 29.826949 bohr reduces the GDF-minus-Ewald difference from -2.210903 to -0.296401 mHa, without entering SAP. This local diagnostic does not propose suppressing that growth in production or establish a converged-SCF discrepancy.

Issue #163 also reports missing hybrid XC output. The renderer now reads result.e_xc when printing BIPOLE’s terminal component history. Four PBE0 RKS/UKS probes, covering converged and capped returns, retain energy -2.837407679083663 Ha and display the stored XC value -0.8198135756723072 Ha. No quadrature is added. Three pure rendering checks cover absent, zero and nonzero XC values. This reporting repair does not resolve the issue’s separate BIPOLE/GDF physical comparison.

The global physical-domain default, all 12 crystal comparisons and the full-node HF/hybrid performance target remain open. The historical numerical archives retain their original source and native identities; a new rebase is not evidence of a fresh execution of those full campaigns.