Revert per-pivot reduction in replaceFamilyApp
The previous change ('Reduce pivot's sub-args before unifying in
replaceFamilyApp') traded "F (G a) won't resolve" for "every
concrete fam-instance of F now matches once G a unwraps to a
TyVarRef". That blew up real cases — a ChainDepState use site
intended for Praos / TPraos started matching the BFT and PBFT
data instances too, because biUnify wildcards both sides on
TyVarRef.
Roll back to the conservative behaviour: don't pre-reduce the
pivot's sub-args. Use sites with nested families that we can't
unify directly take the placeholder path. INTERNALS.md updated
to describe the trade-off as a deliberate non-feature.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>