← board

The repro, four lines of package and two of program

# pkg/__init__.py
VALUE = 27

try:
    from definitely_no_such_module import Image
    have = True
except ImportError:
    have = False



from pkg import VALUE
print("v", VALUE)      # CPython: v 27      pxx: v None

Why it is the expensive shape

Not a crash and not a diagnostic. VALUE is a plain integer constant; it comes back as None because the importer decided the module was unavailable. The module is right there and every other thing about it works — who() called fine in the probe. The trigger is a construct the module wrote about ITSELF, so nothing at the use site or in the importer's own source hints at it.

How it was found, and it is the rule finding itself

Adding a PRESENCE control. The fix in [[bug-n-a-dead-guarded-import-arm-still-compiles-the-module-it-imports]] works by NOT binding a name, and an absence cannot be its own evidence — so the fixture needed a row where the None binding must still happen. Giving the test package a guarded import of its own is what made this fire. No fixture in the suite had a package that guards an import, which is why a bug this loud in effect stayed invisible: the corpus had the construct, the tests did not.

The fix

PyParseImportUnitAs clears SoftUnitMissed before ParseUsesUnit and clears it again afterwards when CompiledUnitCount went UP — i.e. when a unit was actually compiled, so anything the flag now says was said by that unit's own imports. A genuine miss restores the count (pasparser_proc.inc's empty-UnitContent arm under SoftUnitResolve) and is therefore distinguishable without a second flag. An already-compiled unit exits early, writes nothing, and correctly leaves the flag clear.

PyParseOneImport already clears the flag per STATEMENT for the same class one level out — an earlier import's miss leaking into a later one, which cost from tkinter import ttk its binding. That clear cannot reach this case, because here the contaminating write happens DURING our own resolution.

Note

Inert for $(PXX_STABLE) consumers until a pin carries it, and make pin is owner-only.

The fixture, and it took three tries to build one that can fail

test/test_nilpy_a_guarded_import_inside_a_module_does_not_poison_its_importer.npy with test/nilpy_guardpoison/ and test/nilpy_guardpoison_ctl/. Wired into make test-nilpy. Positive control measured 2026-09-11 against 83b883221d70 — HEAD with only the PyParseImportUnitAs hunk removed and rebuilt — where it prints guarded None None against guarded 27 quit.

Three ways a row written for this bug prints PASS while the bug is present, all three measured on that binary rather than reasoned about:

  1. Appended to an existing fixture. The first attempt added rows to test/nilpy_deadarm/, which had already imported that package. The leak fires only when the importer's resolution actually COMPILES the unit; an already-compiled one exits early and cannot be poisoned. Output was byte-identical with the fix disabled. The rows were dropped, not kept as extra coverage — a guard that cannot fail is worse than no guard, because it prints PASS.

  2. try: import ctypes as the guard. ctypes does miss under nilpy — the probe says so — and the fixture still printed 27. The shape that leaks is try: from <absent> import X, which is the shape the ticket's own repro uses and the shape portable packages write.

  3. The control placed beside the subject. This is the expensive one. A clean import resolving ANYWHERE LATER in the file clears the leaked flag before the binding is decided:

    fixture result on 83b883221d70
    guarded, then control 27 — masked
    guarded, print, then control 27 — masked
    control, print, then guarded None — discriminates

    So the control has to come FIRST, which is not where anyone writes one. It is CLAUDE.md's own "a measurement can create the condition it is testing for" in its nastiest position: the contaminant is not an earlier probe step, it is the negative control, the one part of the fixture whose job is to be uninvolved. The fixture says this in its own header so it is not reordered to read more naturally.