Centralised resolver for OpenQP's performance options and the ``perf`` preset.
Each performance knob is an ordinary **input key** (e.g. ``[scf] xc_c2f``,
``[tdhf] resp_cutoff``) backed by a field on the shared ``control`` struct -- there are
no environment variables involved. ``[input] perf = 0|1|2|3`` is a preset that fills in
those input keys (the ones the user left at the sentinel ``auto``) before the config is
pushed to the native library. An explicit input key always overrides the preset.
perf = 0 strict reference / reproducible -- every accelerator off, tightest cutoffs.
perf = 1 recommended production (exact) -- only exact, proven-helpful knobs:
MRSF response cutoff 1e-8, z-vector
warm-start (+ the always-on Fock digest).
perf = 2 faster, tiny degradation -- + coarse-to-fine XC grid + grad Schwarz
cutoff 1e-8 (<=5e-7 a.u. on gradients).
perf = 3 aggressive, degradation allowed -- looser cutoffs traded for speed: grad
1e-7 (~1e-5 a.u.), response 1e-6 (~few ueV).
The accelerators that the CPU benchmark found net-negative or experimental -- Phi-cache,
IncDFT, FP32 and progressive screening (pscreen) -- are NOT enabled by any preset (they
stay available as explicit input keys, e.g. for GPU XC or regimes not covered here).
Precedence (low -> high): control-struct default < perf preset < explicit input key.
An input key set to ``auto`` (its default) defers to the preset; any other value overrides
it. ``perf`` unset (-1) leaves every key at its default -> identical to legacy behaviour.