[wladerer]

Summary

Records the σ_iso(¹⁷O) ecut ladder that was left "pending the run" — proving the shipped matrix-free S-metric CG conduction resolvent (_SMetricResolventCG, response_backend="auto") closes the ecut-divergence of the bare ¹⁷O magnetic shielding on hard first-row anions.

The bare ¹⁷O shielding diverged with plane-wave ecut on the old dense-eigh S-orthonormal conduction resolvent (documented −1912 → −3758 ppm over 40 → 60 Ry; root cause = ill-conditioned augmentation overlap S(k), see analytic_uspp_ecut_divergence.md). The CG fix was shipped but its corrected numbers were never recorded.

Experiment

One fixed MgO ground state per rung, both backends run on the identical SCF + S-metric context (only the resolvent differs). asus @ main e23c776, PAW route (Mg Mg.pbe-n-kjpaw_psl.0.3.0 + committed O O.pbe-n-kjpaw_psl.1.0.0), 2×2×2, symmetry=False.

ecut/rho (Ry)	cond(S)max	dense bare O	dense total O	CG bare O	CG total O
40 / 160	2186	−1871.95	−1519.44	−70.50	+282.01
50 / 200	4233	−3237.23	−2878.09	−72.02	+287.13
60 / 240	5588	−3779.03	−3418.69	−61.80	+298.54
70 / 280	6247	−3395.33	−3035.02	−57.37	+302.95

Verdict — UNBLOCKED

• Dense (old): bare ¹⁷O swings over a ~1900 ppm range with no convergence and a non-monotonic turnover at 70 Ry — the ill-conditioned instability tracking cond(S) (2186 → 6247). The 60 → 70 Ry step alone moves it +384 ppm. Reproduces the documented divergence.
• CG (auto): bare ¹⁷O plateaus (14.6 ppm over the whole 40 → 70 ladder, +4.4 ppm over the top two rungs); total ¹⁷O smoothly convergent. The 40 Ry CG values reproduce the §6 partial record (bare −70.5, total +282.0) exactly.

The bare ¹⁷O ecut divergence is closed by the CG backend. Oxides are unblocked.

Honest caveat (not a divergence): the CG total plateaus near +303 ppm, ~90 ppm above the ≈+215 ppm literature anchor. That residual is stable with ecut and reflects the pseudo choice (non-semicore Mg PAW; the semicore Mg.pbe-spnl of the original report is not on the box) + test-grade settings + frozen-core GIPAW. This PR demonstrates ecut-STABILITY (the blocker), not absolute-anchor agreement.

Contents

• experiments/autoapw/mgo_17o_cg_convergence.py — driver (dense + CG × 4 rungs).
• experiments/autoapw/mgo_17o_cg_convergence.md — writeup + verdict + provenance.
• experiments/autoapw/data/mgo_17o_cg_convergence.json — raw results.
• Cross-reference pointers added to gipaw_absolute_driver_validation.md and analytic_uspp_ecut_divergence.md.

Docs/experiment-only; no src/ change. ~2.95 h asus wall time.

🤖 Generated with Claude Code