vibe-qc 0.7.5 — Fix C (f-shell gradient bug fully resolved)
Patch release on top of v0.7.4 (Whitten's Bridge). Inherits codename.
The direct 4-index ERI gradient kernel
(cpp/src/gradient.cpp::two_electron_gradient_contribution for RHF
and ..._uhf for UHF/UKS) is rewritten as a canonical 1/8 shell-
quartet loop + angular-momentum-canonical reorder before
engine.compute(). The reorder bypasses libint's internal
swap_tbra / swap_tket / swap_braket + DerivMapGenerator path that
produced wrong derivative-to-atom routing on f-shell mixed-l
quartets.
Validation:
* H2CO/def2-tzvp gradient: ~5e-11 Ha/bohr vs PySCF (was ~25 mHa
pre-fix).
* Cs-zero atom 0 c=1 (y): 1e-13 Ha/bohr (was +11 mHa pre-fix).
* Thread determinism: ~1e-15 spread across OMP=1/4/8 (was ~35 mHa
spread pre-fix).
* All 51 gradient tests pass; previously xfail-marked PySCF parity
test now passes by 5 orders of magnitude.
* F atom UHF/def2-tzvp: 4e-13 Ha vs PySCF (machine precision).
* H2CO B3LYP/def2-tzvp: 0.6 µHa diff (grid quadrature, not bug).
The v0.7.3 Fix D auto-route through DF on f-shell bases is removed;
the _auto_df_for_fshell kwarg survives as a no-op for one release of
backwards compatibility (will be removed in v0.8.0).
Known gap deferred to v0.13.5 certification matrix
(HANDOVER_CROSS_CODE_PARITY.md): the DF gradient path still has a
separate ~2 mHa residual on heavy-atom-z components in
cpp/src/df.cpp::compute_3c_eri_gradient_weighted. The direct path is
the rock-solid default again.