Measured 2026-09-13 (frankH), at the tree carrying the module-qualified fix

fec1abd13's sibling: the qualified value read now normalises the def (test_nilpy_a_module_qualified_def_is_a_value*.npy), but only where the scan can SEE the read. Two reductions, four files each, both in one import order and correct in the other.

The return side. retkinds.npy holds def r_three(): return ([1], [2], {"m": 3}); retkreader.npy does Holder(details=retkinds.r_three) and calls it back through the field, exactly as lekkerzeilen's vessel.py:604 does:

import retkinds          # the program imports the DEFS' module first
import retkreader        # ...and the reader second
print(retkreader.three().fittings())

CPython   ([1], [2])
pxx       TypeError: expected a str, list, dict or bytes, got int

Reverse the program's two imports — or drop the first, so the reader pulls the defs in itself — and it is correct. That is what test_nilpy_a_module_qualified_def_is_a_value_across_modules.npy does, and it is written that way ON PURPOSE with this ticket named beside it: a fixture that passes in one order and fails in the other must not be spelled in the order that passes (CLAUDE.md, the ordered-list rule).

The parameter side, which is worse. Same order, annotated parameters:

# ga.py
def g(x: int):   return x + 1
def gs(s: str):  return s + "!"
def gl2(xs: list): return len(xs)
# gb.py
import ga
def run():
    f = ga.g;   print(f(4))        # 5   -- correct
    h = ga.gs;  print(h("a"))      # a!  -- correct
    k = ga.gl2; print(k([1,2,3]))  # SIGSEGV
# d6.py
import ga
import gb
gb.run()

So this is not only a return-convention gap: an unnormalised def also reads its ARGUMENTS through the wrong convention, and a class-typed parameter crashes.

Why a wider scan is the wrong fix

The scan cannot be widened to cover tokens that do not exist yet. A module is lexed when its import is REACHED, so the reads that matter may be lexed after the def is compiled, and no amount of scanning fixes an ordering. Nor may the normalisation be made unconditional: that is bug-nilpy-import-name-forces-function-object-abi, which is recorded as fatal for a Callable[...] parameter.

The shape of the answer

Put the RETURN KIND on the signature record and marshal in the bridge, the way PyHostCall already does for a method (rk := mi^.RetKind, then a thunk chosen by the RESULT with every argument one register wide, and PXXObjRetain on a class result). TPySigRec is {Code, ReqN, TotN, Star, Dflts, Names} and TPyBoundRec carries only IsFunc: Boolean, so today the pair bridge has no way to ask. PYSIG_SIZE is 48 with fields at 0/8/16/24/32/40 — a RetKind at +48 (size 56) and a ParamKinds pointer beside it is the mechanical part; the dispatch is PyHostCall's ptrFamily loop, which is exactly what bug-n-a-star-unpack-through-a-callable-value-stops-at-four-arguments already names as its own non-ladder answer. One trampoline closes both tickets and this one, which is the argument for building it rather than widening anything again.