The JVM supports only unary functions at eval: FuncValue.eval
(values.scala:1042-1056) builds its closure iff args.length == 1 and
otherwise throws "Function must have 1 argument" — at closure CREATION,
so a bound-but-never-applied non-unary lambda rejects too (BlockValue
evaluates ValDefs eagerly). Apply.eval (values.scala:1236-1244) gates
application arity the same way, which sigma-rust already matches.

sigma-rust created the Lambda value for any arity and evaluated
non-unary lambdas to completion. Both serializers are arity-agnostic
(FuncValueSerializer reads a numArgs VLQ), so such trees are reachable
from real wire bytes — an over-accept / chain-split divergence: a tree
carrying a 2-arg (or 0-arg) lambda validates on sigma-rust and fails
on the JVM.

Gate Value::Lambda creation (the single construction site) on
args.len() == 1. The reject is eval-time, NOT parse-time: a non-unary
lambda on the dead branch of a lazy `if` still accepts (pinned by the
vector twin).

Tests drive the four SANTA FuncValue.non_unary_arity vector byte trees
end-to-end (parse → proposition → eval): 2-arg applied, 2-arg bound,
0-arg applied each reject; the lazy-if twin accepts Int 5.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>