Superseded 2026-07-20 by [[feature-nilpy-function-values]]. This ticket was written on the belief that uforth's stored natives are CLOSURES. They are not: 162 of 205 capture nothing, none is defined in a loop, and the few that do capture use Python's explicit _x=x default-argument idiom. The measurement and the correction are in [[decide-nilpy-closure-model]]. Kept for the reasoning about closure records vs cells, which stays valid for whenever real closures are wanted.

SUPERSEDED: nested def as a VALUE (stored, passed, returned)

Split out of [[feature-nilpy-nested-defs]] when its two slices landed 2026-07-20. Nested defs now parse, nest to any depth, and READ enclosing locals — but only when CALLED from inside the enclosing body.

def make(k: int):
    def inner(m: int) -> int:
        return m + k
    return inner            # returning it — not supported
handlers.append(inner)      # storing it — not supported

Why this is the harder half

Capture is implemented by appending the captured values as extra ARGUMENTS at each call site, which only works because every call is lexically inside the parent, where those names are in scope. A def escaping as a value has no such call site: it needs the captured state to travel WITH it — a closure record (code pointer + captured slots) rather than extra parameters.

That is a real feature, not an extension of the current one:

a first-class function VALUE in NilPy at all (Callable[...] annotations exist in PyAnnTypeAt, so the type side has a start);
a heap closure record, and the lifetime rule for it;
Python's late binding: a closure reads the enclosing variable at CALL time, so a loop that defines one closure per iteration and mutates the loop variable has all of them see the LAST value — the famous surprise, and the thing a by-value capture at definition time gets wrong.

Why it matters

uforth registers native words. If that registration stores the inner function in a table rather than calling it inline, the current slice is not enough for the corpus and this ticket is the real blocker. Checked 2026-07-20 with an ast walk over /home/rene/projects/uforth/uforth.py: of the references to inner-def names, 52 are calls and ~197 are value uses (249 name references total, which includes the call's own function reference). So uforth STORES its natives far more often than it calls them inline — this ticket, not the landed slice, is the corpus blocker. Priority raised to match [[feature-nilpy-corpus-uforth]] accordingly.

Gate

test-nilpy green + --tier quick + self-host byte-identical, with a returned closure, a stored-in-a-list closure, and the loop-variable late-binding case, all matching CPython.