Fix: nominal temperature Tnom runs at 0 °C — init param misrouted to hidden_state slot

[Gerd-Vandersteen]

Problem

A model's nominal temperature Tnom can silently run at 0 °C instead of its intended
default (e.g. EKV's DEFAULT_TNOM = 25 °C), mis-scaling every pow(ratioT, BEX) mobility/threshold
temperature correction.

Root cause is a case-insensitive parameter-name collision in the init→eval index map
(vajax/analysis/openvaf_models.py, compile_openvaf_models). A model can expose both an input
param and a same-named hidden_state cache slot that collide when lowercased — EKV exposes input
TNOM (kind=param) and hidden_state Tnom (kind=hidden_state). The map was built last-wins:

eval_name_to_idx = {n.lower(): i for i, n in enumerate(param_names)}

so 'tnom' resolved to the hidden_state slot. Init params are inputs, but the generated init
function then sourced TNOM's value from that hidden_state slot — pre-initialized to 0.0 — so
the model computed Tnom = 0 + P_CELSIUS0 = 273.15 K`.

Fix

Two changes:

1. init→eval map: on a lowercase collision, prefer a real input kind
   (param/param_given/temperature/voltage/implicit_unknown/sysfun) over hidden_state,
   while preserving legacy last-wins among entries of the same preference (so a param and its
   param_given twin are untouched). This is the load-bearing fix — init inputs must never source
   from a hidden_state (init-output) slot.
2. un-given nominal-temp default: tnom/tref/tr fall back to the -\NOT_GIVEN sentinel (model_defaults.get(pname, 1e21)) so the model selects its own DEFAULT_TNOM`.

Evidence

EKV Id at Vg=0.7, Vd=1.0 goes 7.17 → 5.9588 µA, matching an independent VACASK/OSDI run of the
same .va exactly (ratio 1.0000); every d(Id)/dθ matches to 0.00%.

Blast radius (compiled all 11 bundled models): every mapping the fix changes is
hidden_state → input — the same shadowed-input bug — ekv (TNOM), bsimimg (WR, TNOM), bsimcmg
(RHORSD, TNOM, NVTM, CGSP, CGDP), bsimbulk (RBxx, TNOM), hisim2 (mfactor, NFALP1/2, NSUBD/NW).
There are no param ↔ param_given flips. All are latent "init input silently read as 0.0" bugs,
so the promotion is strictly a correctness improvement (these models were not previously
numerically re-validated against VACASK — worth a look during their bring-up).

Note

The cleanest upstream fix for the default half would be teaching openvaf_py get_param_defaults()
to emit sentinel/computed defaults, so the runtime fallback need never guess. The map fix (change 1)
is independent of that and is the essential correctness fix. A companion PR fixes the device
temperature (TEMP); this PR is independent of it.

[Gerd-Vandersteen]

Companion PR: #144 fixes the device temperature (TEMP). Same temperature-defaults family; independent (either can merge alone).

[Gerd-Vandersteen]

Updated (force-push): generalized the collision fix after BSIM4 bring-up surfaced a third case — a model can expose a value param, its param_given flag, and a hidden_state slot all colliding under .lower() (BSIM4 does this for ~176 params: tnom, vth0, u0, ...). The map now ranks collisions value-input > param_given > hidden_state (last-wins within a tier) so an init value-param always reads its value, not the given-flag. EKV and the simple models are unchanged (EKV DC stays VACASK-exact).