mcfastpath.cir

Enhancement-346: random draws on the fast .param path, and the Monte Carlo tier
* Two things, one mechanism.
*
* THE BUG. The fast `.param` sweep only ever re-evaluated brace expressions that
* MENTION the swept name. numparam inlines a `.param`'s expression into the
* device line during preprocessing, so `.param rv = agauss(...)` + `R2 a b {rv}`
* reaches the fast path as `r2 a b {(agauss(...))}` -- a line with no swept name
* in it, which was therefore skipped and left FROZEN, while the reset path
* re-drew it at every point. The fast path silently disagreed with reset.
*
* THE FEATURE. Once random expressions are captured and re-drawn in deck order,
* Monte Carlo needs nothing else: a sample is a re-draw plus an in-place push,
* with no re-source. `montecarlo` now arms the same machinery with no swept knob.
*
* Both rest on consuming the RNG stream exactly as re-sourcing did: random binds
* are evaluated in DECK ORDER, never served from the by-expression cache (two
* devices with identical random text must draw independently), and a sample
* boundary is raised per pass so the Latin-Hypercube sampler steps correctly.
.param rv = agauss(1000, 300, 3)
V1 in 0 dc 1
Rk in m 1k
R2 m 0 {rv}
.control
* the montecarlo command on the fast path. verify_mcfastpath.py runs the same
* experiment as a hand-rolled `reset` loop in a separate invocation -- the path
* this replaces -- and requires the two counts to agree exactly.
montecarlo 200 -analysis op -spec v(m) -max 0.52 -min 0.48 -seed 7
echo SURVIVED
.endc
.end