The Eidos simplifier: occurrence-driven, let-preserving partial evaluation (the retired reduce/specialize/purge loop's successor, doc/eidos.md §8). simplify repeats specialize-on-values, purge, and reduce until every definition is synthable and dictionary-free, bounded by --depth with the retired loop's diagnostic.

• reduceExp is let-preserving: beta redexes become lets, and a let is inlined only when its binder is dead, used once, bound to an atom, or dictionary-typed (dictTyCons — never representable, so substituted at every occurrence); a representable multi-use binding KEEPS its let (sharing). The one lambda-lifting policy surviving from the retired pipeline is explicit here (Clash's LiftNonRep): a multi-use *function-typed* let cannot stay (it is not representable), so it is lifted to a top-level definition over its captured locals, named with the $LL. compiler-lifted prefix (display-only: the machine fold strips it when naming the Hyle definition). Top-level references are never unfolded — the function hierarchy is preserved; higher-orderness dies by argument baking, not call inlining.
• specialize on values (the retired value-argument specializer): a call to a top-level definition with *closed* arguments (free variables all top-level) bakes those arguments into a memoized clone named from the baked-argument canon (originTag). Self-calls in the clone still reference the origin — the memo closes recursive loops. Memo keys are alpha-canonical: arguments are renumbered from a disjoint unique supply and printed. Baking is also how non-representable scalar arguments (Integer, String) disappear: their call sites pass closed values.
• purge: definitions unreachable from the device root (plus the builtin signature carriers, which carry the builtins' type assumptions to the Eidos-to-Hyle fold).

Join points are preserved (dead ones are dropped); datatypes are left untouched.