[krystophny]

Part of #7; stacks on #10. Adds the raw radial-normalized partial projectors, density matrices and local spinor rotations written to case.parproj for the non-mixing SP+SO path.

The generator keeps separate data for every (atom, l) shell, validates matching spin windows and uses each spin's own band offset. Unsupported projection modes, non-SO inputs, cubic p/f bases and nonidentity struct rotations fail explicitly pending their later port steps.

Validation: native CaOs2 partial-projector fixture agrees within 1e-7 with exact integer fields and finite floating-point comparisons. An independent identity-symmetry regression preserves distinct s and p density blocks on the same atom. Local NumPy tests pass; full TRIQS integration runs in CI. The tighter double-precision comparison belongs to #12 with #14.

Chris&AI

[krystophny]

Accuracy: reproduces the released dmftproj case.parproj to 1.2e-8; 1.5e-14 (machine precision) with the precision fix #14 + exact Python #12 (well-conditioned radial sqrt, no amplification).

[krystophny]

GPT Sol rebased this onto the reviewed #10 and completed the remaining fixes: shell-specific projector/density tables, native window selection, spin-window validation and per-spin offsets. Added a two-shell identity-symmetry regression and tightened the native comparison to finite values within 1e-7. Local tests pass; CI is running.

Chris&AI

[krystophny]

GPT Sol fixed local-rotation parsing when atomic sorts outnumber symmetry operations, corrected per-shell representation parsing, and rejected unsupported multiple local orbitals. The P1 regression and all four preceding generator tests pass. Restacked on updated #10. Latest-head CI is running/queued.

Chris&AI

[the-hampel]

Rebased onto unstable now that #10 is in. Your two commits are unchanged (range-diff identical); new head 7926555.

Checked locally:

• full ctest 27/27
• CaOs2_partial.parproj.gz matches native dmftproj -so -sp on the same inputs (max diff 2.2e-16, last-digit noise)
• Python generator vs native: max abs diff 1.2e-8 (test tolerance 1e-7); perturbing a reference value by 1e-6 in the projector or the density-matrix block makes the test fail

Looks good to us. Non-blocking points:

1. _euler is never called.
2. Unreachable non-SO paths, as in [wien2k] pure-Python case.sympar partial-shell symmetry matrices #10: write_parproj requires SO and both almblm spin files, so the single-file branch, ns = 2 if ifSP else 1 and the non-SO branch of _srot_rotrep_nonmixing cannot run. _symmetrize_densmat takes info, ns, ifSO without using them.
3. The module docstring calls the density matrix "tetrahedron-weighted", but the code point-integrates with tetr[0] as k-weight. Line 347 has an empty section header.
4. Coverage gaps, just noting: for all three CaOs2 shells the first op mapping the representative atom is unitary, so the time-reversal branches of the Rloc / rot_projectmat path are not exercised; and the single k-point fixture does not test the k-weighting.
5. The Wigner-D / indmftpr / dmftsym / cubic-basis helpers are now duplicated across symqmc, ctqmcout, sympar and parproj; best consolidated in [wien2k] complete the port to machine precision: shared module, fromfile, proj_mode, outband #12.

If you don't get to 1-3, we'll add a small cleanup commit ourselves and squash-merge once CI is green.

---

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