perf(solvers): zero-copy col-major QR native kernel + call-time GRADWAVE_EIGENSOLVER read

[wladerer]

What

Two changes to the native Davidson stack:

1. Kernel cleanup (solvers/native/davidson_native.c) — zero-copy col-major QR (zgeqrf/zungqr run directly on the basis buffers, exploiting that a row-major (rows, m) block is byte-identical to the col-major (m, rows) matrix whose columns are the basis vectors — deletes 2 explicit + ~4 hidden LAPACKE transpose passes per orthonormalization), planar-double SIMD elementwise loops (kinetic, v_eff multiply, residual norms, Teter directions), and sphere-only scatter-buffer zeroing instead of a full (n+1)-point memset per band (m/n ≈ 0.06–0.1).
2. GRADWAVE_EIGENSOLVER read at call time, not import time (scf/loop.py) — the import-frozen copy silently ignored os.environ set after import gradwave. This is not cosmetic: it silently degraded every "native" arm of a benchmark ladder to eager (identical wall to eager at every thread count was the tell). Tests now pin the override via monkeypatch.setenv to lock in the call-time contract.

Measured (asus, 22-core, same-harness A/B, per-solve spy verifying the native path engaged — zero fallbacks; identical F and iteration counts in every arm)

Al-4 conv 4³ FD and Al-1 fcc 8³ FD, ecut 30 Ry, full SCF wall:

arm	eager	old kernel (#467)	this PR
Al-4 1-thread (batch)	93.0 s	54.0 s	53.7 s
Al-4 1-thread (retire)	—	34.0 s	34.7 s
Al-4 8-thread (retire)	—	7.7 s	7.5 s
Al-4 default threads	34.0 s	7.8 s	7.7 s
Al-1 1-thread (batch)	53.3 s	21.7 s	21.7 s
Al-1 1-thread (retire)	—	18.0 s	17.8 s
Al-1 8-thread (retire)	—	4.3 s	4.2 s
Al-1 default threads	19.8 s	4.4 s	4.2 s

Honest verdict: the kernel rewrite is an end-to-end wash vs the old kernel (≤2%, within noise). It ships because it is strictly cleaner (fewer memory passes, no LAPACKE hidden copies), never slower, and the measurement campaign it enabled corrects two records:

• The probe-era claim that the native kernel had a 0.32–0.45× single-thread deficit does not hold end-to-end: the old kernel was already 1.7–2.5× eager per thread e2e (93.0→54.0 s, 53.3→21.7 s). The kernel-level probe number was apples-to-oranges with e2e wall.
• The corrected single-process native headline (this harness): Al-4 34.0→7.5 s (4.5×), Al-1 19.8→4.2 s (4.7×) with one knob (eigensolver: davidson-native), identical free energies to all printed digits. Last night's "9.8 s" figure was harness-inflated.

External proxy note: VASP is not runnable on this infrastructure; QE (same box, same ONCV pseudos/ecut/k-mesh) remains the honest external reference — QE pw.x Al-4 best 0.6–2.3 s, so the small-cell gap is now ~3–8× (was ~2–9×).

8/8 native unit tests pass against the new .so; the fast tier passes on asus with the branch .so active.

🤖 Generated with Claude Code

https://claude.ai/code/session_018j6BBnPtBSjkrnASSc52mT