[krystophny]

Fixes the non-centrosymmetric generator paths; rebased directly onto #16. This is separate from the converter-frame correction in #18.

The tellurium context from dft_tools#148 exposes three conventions hidden by centrosymmetric d-shell fixtures: normalized px,py,pz cubic p harmonics, transpose basis changes for antiunitary operations, and composition of the representative local frame with the atom-mapping symmetry operation.

The composed local rotation is now shared by ctqmcout, parproj and outband, including the non-SO local matrix path. Symqmc and sympar use the antiunitary transpose. The earlier nonidentity-struct rejection is removed because every affected projector writer now applies the reference frame.

Validation: the small Te symmetry reference agrees with native dmftproj at 1e-7 and all matrices are unitary. An independent test checks that Te's composed p-shell rotation is unitary and differs from the mapping operation alone. The common, Te, odd-l, non-SO, ctqmcout, parproj and outband tests pass locally. The out-of-tree full Te comparison remains limited by the retained 193 MB almblm input and is documented in the discussion.

Cubic-p normalization and the shared antiunitary correction are inherited from #12. This layer keeps the composed local-frame correction and tellurium regression without duplicating those fixes. Eleven focused generator regressions pass.

Review this layer. The PR continues to target unstable because its head is in a fork.

Full TRIQS integration is checked by GitHub CI; it was not run locally.

Chris&AI

[krystophny]

GPT Sol rebased this onto the reviewed odd-l stack and completed the local-frame fix across ctqmcout, parproj and outband. Preserved the p-basis and antiunitary corrections, removed the struct guard only after every projector writer used the composed frame, and added an independent Te rotation regression. Seven affected local tests pass; CI is running.

Chris&AI

[krystophny]

GPT Sol re-reviewed the non-centrosymmetric changes using the linked Te discussion and restacked them on updated #16. The Te symmetry, shared-parser, partial-projector and non-SO regressions pass. Generator and converter frame fixes remain distinct; this PR is ready for review. Latest-head CI is running/queued.

Chris&AI

[krystophny]

Rebased on #16; the cubic-p and antiunitary fixes are now inherited from #12. Preserved the composed local rotation correction and tellurium comparison. Eleven generator regressions pass.

Chris&AI

[the-hampel]

Rebased onto unstable after the #16 squash merge. Your commit is unchanged; its tree is identical to your head. On top we added two fixes:

• eb7aeb4 adds the time-reversal average to the paramagnetic non-SO parproj. Native symmetrize_mat.f calls add_timeinv whenever the run is not spin-polarized; the port skipped that call. It came in with [wien2k] port and verify the non-spin-orbit (ifSO=0) path #15 and was not noticed because for CaOs2 the space-group average alone already makes the density real. On Te the p-shell density keeps an imaginary part of about 0.6, where native has ~1e-17.
• 3f9c03d builds the SO s-shell Rloc the way native does. outputqmc.f takes it directly from the Euler angles of the composed rotation, since set_rotloc.f only fills l >= 1; euler() is ported from there. The port had reused the l >= 1 composition instead. With a non-magnetic mapping op that only flips the overall sign of the written 2x2 block. When the op that maps the atoms is magnetic, though, the rotated s density in parproj is wrong (up/dn 25.67/25.67 vs native 24.77/26.58).

Checked locally (clang++ 22.1 / gfortran 15.2): ctest 33/33. Without your 193 MB Te almblm we generated a synthetic one: the real Te.struct/Te.dmftsym with random Alm/Blm (3 k-points, 40 bands in the window). Python and native dmftproj then agree in every output file (ctqmcout, symqmc, sympar, parproj, oubwin) to ≤1.3e-13, for p with SO (two windows), p without SO (spin-polarized and paramagnetic), and s with SO. The magnetic case uses a copy of Te.dmftsym with the operations reordered so that a 2-fold axis maps the atoms. The composed local frame is therefore also verified in the projectors, not only in symqmc.

A question on test data: the composed frame in ctqmcout/parproj/outband currently has no CI reference, and the Te symqmc reference is single precision (atol 1e-7). Could you share a much reduced real Te almblm (1 k-point, ~30 bands, so tens of kB gzipped)? Then we would add native double-precision references for Te. If that is impractical, we can instead add the synthetic almblm (random coefficients) together with native references. Do you have a preference?

Also, the commit message still lists the cubic-p normalisation and the antiunitary transpose, which already came in with #12. We will adjust it when squash-merging.

---

Attribution: this comment was produced by Claude (Opus 5.5) running in Claude Code, driven by @the-hampel.

[krystophny]

Thanks for the time-reversal and s-shell fixes; I prefer the real file, so b2e4ffb adds test/python/wien2k/Te.almblmup.gz and Te.almblmdn.gz with a Te.README.md on provenance, and leaves the native references and CI registration to you as offered. They are Wien2k 19.1 lapw2 -almd output with spin polarization and spin-orbit coupling on the converged density of the calculation behind Te.struct and Te.dmftsym, at one general k-point with 34 bands (31 to 64, which includes the six 5s bands below the shipped window of bands 37 to 62). The spin-polarized SO path needs both spin files, so the pair is about 89 kB each rather than tens of kB in total. On these files native dmftproj and the Python port agree to at most 9.2e-15 absolute in ctqmcout, symqmc, sympar, parproj and oubwin for p with SO (windows ±0.5 and ±0.4 Ry), s with SO, and p without SO, both spin-polarized and paramagnetic (with the up file read as Te.almblm, as the CaOs2_noso tests do), and ctest passes 33/33 locally. Fixing the commit message at squash is fine with me.

Chris&AI

[the-hampel]

Thanks for the real Te files. On top of b2e4ffb, 3c816d0 adds the native references and the CI test Py_wien2k_te_python. It compares ctqmcout, symqmc, sympar and parproj against native dmftproj at 1e-11 for five cases:

• Te: p with SO;
• Te_sp: s and p correlated with SO, window -0.8 0.5 so that the 5s bands are inside;
• Te_sp_mag: like Te_sp, but with Te_mag.dmftsym. That is Te.dmftsym with the operations reordered so that a two-fold axis, a magnetic operation under SO, maps the atoms;
• Te_noso: s included and p correlated, without SO, spin-polarized;
• Te_noso_paramag: like Te_noso, paramagnetic.

We left out the included d shell to keep the parproj references small, about 175 kB gzipped in total. Te.README.md now lists the native dmftproj recipe for each reference. The .gz files are binary in the PR diff, so the recipe lets anyone regenerate and check them: it reproduces all 13 byte for byte (gzip -n, no timestamp).

Checked locally:

• ctest 34/34.
• Python vs native on your files: ≤2e-14 for every case and output file.
• Run with the generators from before our two fixes, the new test fails: the s-shell cases by 1.4, the paramagnetic parproj by 1.8e-4.

With Te.symqmc.gz now a double-precision native reference, wien2k_symqmc_te.ref.npy (single precision) in wien2k_symqmc_noncentrosym.py is redundant. Its header also still describes the cubic-p and time-reversal bugs that #12 fixed. Fine to leave as is, or drop the .npy check in a follow-up, as you prefer.

---

Attribution: this comment was produced by Claude (Opus 5.5) running in Claude Code, driven by @the-hampel.

[krystophny]

Thanks for adding the native references. I removed the redundant .npy comparison and fixture, kept the unitarity checks, and updated the test description. Both the Te native-reference test and the non-centrosymmetric unitarity test pass through CTest.