← board

The last class in a module reads every attribute as zero

This ticket was corrected once. The first version said "a class-only module", and said any module-level statement fixed it. Both were wrong; the history is in ## What the first version got wrong at the bottom, because the way it was wrong is a reusable warning about measuring.

Repro

nodemod.py:

class Node:
    ELEMENT_NODE = 1
    TEXT_NODE = 3
    DOCUMENT_NODE = 9

main.npy:

from nodemod import Node
print(Node.ELEMENT_NODE, Node.TEXT_NODE, Node.DOCUMENT_NODE)



pxx:     0 0 0
CPython: 1 3 9

Strings degrade the same way (S = "hi" reads back empty), so the attribute is being read at its type's zero value — the storage exists and was never initialised.

The rule, measured

A class whose definition is not followed by a module-level statement loses its class-attribute initialisers. It is positional and per class, not per module:

class K:
    A = 7        # K.A reads 7   -- a module-level statement follows K
TOP = 1
class J:
    A = 9        # J.A reads 0   -- nothing follows J



pxx:     K.A= 7   J.A= 0
CPython: K.A= 7   J.A= 9

So the shape that gets hit is the last class in a module, which is an entirely ordinary way to end a file.

imported module's body mod.K.A
class K: A = 7 (nothing else) 0
class K: A = 7 then TOP = 1 7
TOP = 1 then class K: A = 7 0
class K: A = 7 then print("init") 7
print("init") then class K: A = 7 0
class K + TOP = 1 + class J K=7, J=0

Before does not help; only after. The statement kind does not matter (an assignment and a print behave identically) — position does.

Not the discriminator, each tested and each ruled out: the class name, the attribute name, the import form (import m / from m import K / both / through an instance / dotted import a.b), and shim-vs-plain-module. The same class defined in the reading file is always correct — which is why this survives local testing and only appears once the code moves into a module.

Methods are unaffected: a class with only defs and no class-level attributes works fine, whatever follows it. Only attribute initialisers are lost.

Reading of the cause — NOT verified, do not bank it

Position-sensitivity that is fixed by a following statement suggests the class-body assignments are emitted into module init but the init routine is truncated at, or generated only up to, the last module-level statement — so initialiser work attributed to a trailing class falls off the end.

This is a story, not a measurement. It was not checked against PXXDBG=a.ir: or the emitted init. Whoever takes this should print what the compiler actually emits before trusting the paragraph above — and note that the first version of this ticket also had a confident-sounding cause paragraph built on a boundary table that was partly wrong.

Why 60

Not 70+ only because no shipped code is currently wrong (see below). That is luck, and it expires the first time someone writes the obvious constants shim.

Exposure in our own shims — the corrected answer

The first version of this ticket claimed lib/rtl/mimic_six.py and lib/rtl/mimic_warnings.py "escape by accident because they carry module-level assignments". That reasoning does not survive the corrected rule. Re-checked, both are safe, for two different reasons and neither is the one first given:

So the exposure is real but currently nil, and the guard to remember when writing any future .py shim is: a class with class-level constants must not be the last thing in the file until this is fixed. Both shims' tests pass and would keep passing either way, so review cannot catch it.

Gate

The repro prints 1 3 9. Every row of the rule table above matches CPython, including J.A = 9. Worth a regression test in the dual-runnable form (decide-what-an-unwired-test-may-assert) — the module and the reader are both legal CPython, so the oracle stays attached to the test rather than being consulted once.

What the first version got wrong

Two claims, one root cause in how they were measured.

  1. "A class-only module" — wrong scope. It is per-class and positional, so a module with several classes silently corrupts only the trailing one.
  2. "Any module-level statement fixes it" — wrong condition. Only a statement after the class does.

Claim 2 came from a probe whose output was piped through head -1. In the two rows where the module-level statement was a print, the print's own output was the first line, so what got recorded as "the attribute is correct" was really "the print ran" — the attribute value was never in the captured output at all. The rows were not measured; they were mislabelled.

The coordinator caught it by running the control the table predicted — an assignment above the class — getting 0 0 0, and reporting the disagreement instead of assuming its own setup was broken. Worth stating plainly: the harness that formats a measurement can silently answer a different question than the one asked, and a table row is only as good as the bytes it actually read. head -1 on a probe whose output length varies per row is now a known trap.

The one conclusion downstream of the wrong row — "our shims escape because they carry module-level assignments" — was also wrong, and is corrected above. It mattered: it was load-bearing for "no shipped code is affected".

FIXED 2026-08-18 (frank2-7e) — the hoist queue was never drained at end of module

Measured, not reasoned — and the ticket's own cause paragraph was wrong

The unverified reading above ("the init routine is truncated at the last module-level statement") is not what happens. PXXDBG=a.ir:* on the two-class repro shows __init_nodemod in full:

0: const_int ival=7
1: store_sym [sym=$clsattr.K.A]
...
12: call  (pyclsattr_bind for K.A)
13: const_int ival=1
14: store_sym [sym=TOP]

Nothing is truncated: TOP — the LAST module-level statement — is present and correct. J is simply absent entirely. The routine is complete; the trailing class's work never entered it.

Root cause

PyEmitClsAttrBinds publishes each class attribute with PyHoistStmt, i.e. onto the hoist queue. In ParsePyUnit's module loop only the ordinary-statement branch drains it (stmt := PySeqAppend(PyFlushHoist(-1), stmt)); the class and def branches flush PyFlushDefInit and never PyFlushHoist. After the loop there was no final drain at all, so whatever the last construct hoisted was dropped on the floor before CompileAST(seqNode).

That is exactly the measured boundary: only a statement AFTER helps, and its kind does not matter — any ordinary statement flushes the queue, which is why an assignment and a print behaved identically, and why a statement BEFORE does nothing.

Fixed on BOTH arms — the program loop has the same hole

ParsePyProgram's loop is built the same way and also had no final drain. There the bug is masked rather than absent: any use of a class is itself a following statement, so a program that reads the attribute has already flushed the queue. That is the whole reason "the same class defined in the reading file is always correct", which the ticket recorded as a fact without a cause.

Both loops now drain after the loop. Fixing only the observable arm would have left one defect reachable through two paths, with the unobservable one still broken — devdocs/dev/normalise-dont-special-case.md, and the reason that file says to grep for the sibling before closing.

Verified by value against CPython

shape before after CPython
class LAST in module, 3 int attrs 0 0 0 1 3 9 1 3 9
a str attribute on a trailing class '' hi hi
trailing class read through an INSTANCE 0 1 1
class K / TOP=1 / class J K=7, J=0 K=7, J=9 K=7, J=9
statement after the class (was already fine) 7 7 7

Regression test

test/test_nilpy_last_class_in_module_attrs.npy + test/nilpy_units/lastclassmod.npy, wired into test-nilpy by name.

The module deliberately contains an Early class with statements after it as well as a trailing Node: on a broken compiler Early is correct and Node is not, so the test pins the positional rule rather than just "class attributes work". Verified both ways — on pinned v348 it prints 1 2 / 7 / 0 0 0 / (blank) / 0; at HEAD and under CPython it prints 1 2 / 7 / 1 3 9 / node / 1.

Gate

make compiler/pascal26 (fixedpoint, converged) + the value table + the new test failing pre-fix and passing post-fix + tools/gate.sh quick GREEN. No pin needed; nothing in compiler/builtin/**.

Log