stdaudit_examples

Standard-analyses audit: referee battery + three fixes (Enhancement-179)

The gap-analysis doc marks the "Standard analyses (analog)" table all on-par — but several rows are 1990s SPICE3 code whose values had never been checked against independent physics (the E-171/175/177/178 accidental-correctness lesson). This suite embeds the referees. Both solvers, every check.

The three fixes

1. .tf current-output impedance (tfanal.c): .tf i(vm) vin always reported output impedance 1e20 — the code clamps 1/MAX(1e-20, rhs) where the branch current of the unit forcing is negative for every passive network (the input-impedance path right above correctly divides by -rhs). Inherited verbatim from Berkeley SPICE3 — a 35-year-old bug. Now exact (RL + node Thevenin, digit-perfect vs hand analysis).
2. KLU AC sensitivity truncation (cktsens.c): the KLU complex-conversion block inside the frequency loop reused i — the outer loop variable — so after the first point i = DEVmaxnum ended the sweep: one frequency point, silently. The surviving point was correct, which is why E-62's single-point check passed. Now the full sweep matches the analytic dV/dC = −jωR/(1+jωRC)² to 6 digits under both solvers.
3. .meas DERIV{ATIVE} implemented (com_measure2.c): parsed since the SPICE3 era but never evaluated — an explicit "currently not supported" stub with an empty #if 0 measure_deriv() placeholder. Implemented as a 3-point Lagrange-quadratic derivative on the nonuniform time grid (AT= and WHEN forms), verified against the analytic sine slope to 5 digits. The DERIVATIVE/INTEGRAL long spellings are now accepted too.

Measured-correct (the rest of the table)

• .disto — the 1990 Volterra code holds up impressively: HD2/HD3 match an analytic diode-kernel referee with a frequency-dependent load (harmonic loads correctly at Z(2ω)/Z(3ω) — the E-177-style frequency probe is clean) to ≤0.06%; the SIM2 two-tone path (f1+f2, f1−f2, 2f1−f2) matches including cascade terms; Volterra amplitude scaling is exact; and nonlinear-junction-capacitance harmonics agree with the E-134 Harmonic Balance engine to ~6 digits — two fully independent engines.
• .noise integrals — onoise_total² equals the band-limited analytic (→ kT/C) to 0.06% and the flicker log-integral to 6 digits.
• .sens — DC sensitivities at a nonlinear OP (dv/dRs, dv/dIS) match central finite differences to 5–6 digits, model parameters included.
• .pz at a nonlinear OP works — the E-62 "nonlinear pz quirk" was the input convention (a bias source on the injection node shorts it; ngspice's refusal is correct). The driving-point form returns the linearized pole −(1/Rs+g_d)/C to 0.02%.
• .tran/.op/.meas — trap/gear both ≤4e-7 vs the analytic RC decay; hard-DC homotopy identical across solvers; RMS/PP/INTEG/WHEN ≤1e-6.

Running

python3 verify_stdaudit.py     # 8 checks x {sparse, klu}
python3 make_stdaudit_fig.py   # figure