inputguard_examples

inputguard_examples — Enhancement-426

Inputs ngspice accepted without ever checking them, and the confident wrong answers that followed.

python3 verify_inputguard.py

88 checks. Every one is measured as a number, and every boundary is pinned from both sides — the refusals and the legitimate spellings that must keep working. That second half is the point of the suite: three proposed fixes were killed during review because they would have rejected something real.

	what was accepted	what it produced
1	a mistyped output node, in the session's first analysis	tf v(a,nosuch) = 1.0; sens all -0.0; onoise_total = 0.0 — and a heap read at 3.999110e+252
2	ac dec 10 100k 1k	a different sweep: 31 points over 1e5…1e8
2	noise with an inverted range	onoise_total = 0.0, from a loop that never ran
2	.dc with a step pointing away from stop	a plot containing no vector
3	meas MAX @dev[param] without a .save	the one-point snapshot, reported as the maximum
4	temp = -500	a negative thermal voltage, $vt = -0.0195 V
5	itl2 = 0	736,920 Newton iterations instead of 55
6	m = -1	a 2 kΩ resistor stamping −2000 Ω
7	1e400	inf — a resistor silently became an open circuit
8	a failing OSDI setup_model followed by a good one	SIGSEGV, zero bytes of output

What must keep working, and is checked here

• equal sweep endpoints — 19 .ac, 4 .noise, 9 .sp cards in this repo
• dc v1 1 1 1 (single point, 13 decks) and dc v1 2 0 -0.5 (descending, 2 decks)
• m=0 — the ordinary "disable this instance" idiom, deliberately left silent
• itl6=0 — a table synonym for srcsteps, documented, 4 decks rely on it
• temp=-25 — ordinary; the line is absolute zero, not freezing
• 1k2, 2meg5, 1e5x, 1kk, 0x10, 5kohms — documented scale-factor forms
• a .tf card placed before the devices that define its nodes (E-349's case)
• @r1[resistance] still returning the nominal value it was given

Note on the harness

run() captures stdout and stderr together, deliberately. ngspice writes its own $finish/$stop Notes to stdout while the OSDI log callback writes WARN/ERR/FATAL to stderr — a check that watches one stream scores the other as silent, which is exactly how the original report came to claim $finish printed nothing.