Enhancement-644: a paramset's own parameters are instance parameters, and an instance selects its member (IHP hunt C1, C2)

Enhancement-644: a paramset's own parameters are instance parameters, and an instance selects its member (IHP hunt C1, C2)

Scope: C1 and C2 of the
2026-09-15 IHP SG13G2 hunt.
Compiler: openvaf/sim_back/src/module_info.rs (a paramset's own parameter
is is_instance unless (* type="model" *)). ngspice:
src/osdi/osdiinit.c (osdi_select_paramset_member with an instance line's
parameters; osdi_paramset_family_head/_of; the collision warning skips a
bound parameter; no bare i alias over a model's own i),
src/spicelib/parser/inp2n.c (the per-instance selection: the first n
line binds the card, a later one needing another member gets a clone card),
src/spicelib/parser/inpgmod.c + inpmkmod.c + include/ngspice/inpdefs.h
(INPmodOsdiSel), include/ngspice/osdiitf.h. New suite
paramsetinst_examples (14 checks per
solver). Both sides.

Suites: paramsetinst_examples 14 of 14 per solver, both solvers (7 fail
on the E-643 binaries); every paramset suite (paramset, paramsethsp,
paramsetlrm, paramsetoverload, paramsetguard, modelparamset,
cardbound, macrofold, …) green; workspace cargo test green; full sweep
517 of 517.

What was wrong

C1. A paramset's own parameters became the twin module's model
parameters, so the one way to set them was a .model card:

n1 n1 0 0 rsil l=0.5u w=0.5u mm_ok=0
Error … unknown parameter (l): it is a model parameter of this device -- set it on the .model card, not on the instance line

That is not what a paramset is. LRM 6.4 makes the paramset's parameters
what an instance of it sets — rsil #(.l(0.5u), .w(0.5u), .mm_ok(1)) r1(n1, gnd, gnd) in the IHP decks — and 6.4.2 selects the member per
instance from exactly those values. A SPICE device library is written the
same way: one card per device kind, geometry on the instance line. With
model-only parameters every geometry needed its own .model card, and the
selection could only look at the card.

C2. The E-335 case-collision warning ("declared more than once differing
only in case; only one of them can be set from a netlist") fired for pairs
in which nothing can be set wrongly: a target-module parameter the paramset
bound — PSP's SWSOA under the paramset's own swsoa, sp_resistor's
r alias under the paramset's R — is PARA_FLAG_FIXED and takes no
netlist value at all; and the loader's own bare i alias for a two-terminal
device (E-394) collided with a model's (* desc="Current" *) real i, where
the model's entry wins the lookup anyway.

What changed

• Compiler. A parameter declared inside the paramset a twin was
  lowered from (Module::paramset_decl_range, E-643) is an instance
  parameter unless it carries (* type="model" *); a pass-through
  target-module parameter is what its own declaration says. Everything
  downstream already knew how to treat an instance parameter: the card
  gives the instances' defaults (showmod lists them as "instance defaults
  on this card"), the instance line and alter set them, and E-546's
  fixpoint promotes the bound parameters whose overrides read them
  (.resistance = 7.0 * l / w) to per-instance evaluation, silently, as it
  does for any localparam.
• Selection per instance. osdi_select_paramset_member(type, card, inst, …) takes the instance line's name=value pairs beside the card's
  — an instance value replaces the card's for the same name; a name no
  member declares (m=, temp=, a misspelling INPdevParse refuses later)
  takes no part. In INP2N, before the card is materialised: a card that
  heads a family and is not yet bound is bound to the member this instance
  selects (INPmodOsdiSel records it, so create_model does not select
  again); a card already bound to another member gets a clone,
  <card>.<member> (the bin cards' naming), made once by INPmakeMod on
  the same card line and shared by every instance selecting that member —
  with a note: n5: paramset 'rsil' resolved to its member 'rsil__2' for this instance's parameters, where .model rsil is bound to 'rsil' -- a second card, rsil.rsil__2, is made for it (LRM 6.4.2). The card-only
  route (E-565, a card used by no n line yet) is unchanged, and an
  instance that gives nothing on a card that gives nothing is still the tie
  error.
• C2. osdi_warn_case_collisions takes the descriptor and skips a pair
  in which either entry is PARA_FLAG_FIXED; the bare i alias is not
  synthesised when the model declares an i of its own (instance parameter
  or operating-point variable), so @n1[i] is the model's value as before
  and nothing is warned. Two settable gain/GAIN still warn.

Verification

check	result
n1 1 0 r0 l=1u w=0.5u mm_ok=0 over a paramset with own w, l, mm_ok and a (* type="model" *) tc	accepted, read back, I = 1 V / 14 Ω; tc on the instance line refused as a model parameter; the card's l=2u reaches an instance that does not set it (1 V / 28 Ω); showmod lists w under the instance defaults and tc among the model parameters
.model r0 rs + n1 … mm_ok=0 + n2 … mm_ok=1	n1 is rs (7 Ω), n2 is rs__2 (mm = 2, 3.5 Ω) through the clone r0.rs__2, noted
the first instance selects rs__2; two later mm_ok=0 instances	the card is bound to rs__2 (noted), n2 and n3 share one clone r0.rs
card mm_ok=0, first instance mm_ok=1	the instance's value selects (rs__2), and being first it binds the card itself
mm_ok=2 on the instance / m=2 temp=30 beside mm_ok=1 / nothing anywhere	refused with each member's reason / selected, the builtins ignored / the tie error
IHP resistor_paramset.va (E-641's fold, E-643's ranges): one .model rsil rsil card, l/w/mm_ok on the instance lines	rsil/rppd/rhigh at 1 mA give the gnucap values; an mm_ok=1 instance beside mm_ok=0 ones gets rsil.rsil__2
IHP sg13g2_moslv_paramset.va: one .model nmos sg13g2_lv_nmos_psp card, w, l, ng, as, … on two instance lines	Id = 85.558 µA at w = 0.3 µ (the gnucap reference) and 174.9 µA at 0.6 µ, from one card
the collision warnings on the IHP objects (sp_resistor's i; ptap1/ntap1/Rparasitic's r; PSP's swsoa)	none; gain/GAIN still warns
every paramset suite; cardbound, macrofold	green
workspace cargo test	green
paramsetinst_examples	14 / 14 per solver, both solvers; 7 / 14 fail on the E-643 binaries
full sweep	517 of 517

What this does not do

alter of an own parameter after the instance is bound (alter @n5[mm_ok]=0)
changes the value, not the member — selection is an elaboration-time act in
the LRM too, and the member's own range (mm_ok from [1:1]) refuses the
value. The clone card's name (rsil.rsil__2) is an internal one, listed by
showmod like a bin card's.

Write-up: enhancements_doc/Enhancement-644.md