The Java backend printed local MLfix as a statement sequence
(var f = ...; ...; f.apply(args)), which is invalid in expression
positions (lambda bodies, let RHS, function arguments). Every other
MiniML construct already prints as an expression.

Encode a local fix applied to K arguments as a call to a generated
generic helper fixK(gen, a1, ..., aK) that ties recursion by method
self-reference, so no cell arrays, unchecked casts, definite-assignment
or effectively-final issues arise. Argument types ground A1..AK
bottom-up and R is grounded by the target type of the expression
position. Helpers are emitted on demand per arity next to the let
helper (fix0 for unapplied fixes implies fix1).

Mutual fixpoint groups are eliminated before printing by a Bekić
rewrite (unmutualize) into nested single fixes, using mlutil's
permut_rels and gen_subst (now exported).

Supporting changes needed for javac type inference:
- Inline let-bound fix heads into their use sites (a fixK call in the
  targetless let() argument position would infer Object).
- Upcast constructor expressions to their inductive interface,
  ((natlist) new Cons(...)), to avoid over-specialized inference such
  as A2 := Nil; skipped for custom inductives. Expected outputs
  regenerated accordingly.
- Fix MLmagic and MLcase dropping the application args accumulator.

Known limitations: fixK calls in positions without a target type infer
R = Object (e.g. a fix application result bound by let); helper names
fixN can collide with user identifiers (same caveat as let);
parameterized inductives get raw-type casts.

Tests: local_fix (single recursion, applied K=2 and reused let-bound
fix) and local_fix_mutual (mutual even/odd, K=0 elim-last and K=1
elim-first paths), each with a runtime driver.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>