← board

from collections.abc import ... is swallowed by the collections root rule

Filed 2026-08-19 from [[feature-b-mimic-collections-abc-mapping-and-mutablemapping]]. This ticket, not the shim, is what unblocks the corpus.

Measured on pinned v356 (2bb09afb0cff):

from collections.abc import Mapping
print(Mapping)
CPython <class 'collections.abc.Mapping'>
pxx (pinned v356) error: undefined variable (Mapping)

import collections.abc as cabc does reach the shim (the resolver prints note: collections_abc -> mimic_collections_abc (shim, subset)), so the shim itself is fine — it is the from-form that never gets there.

Root cause

PyImportRootIsConsumedOnly (compiler/pyparser.inc:33003) decides whether a from-import should be consumed and ignored, and it tests only the root of the dotted name. collections is on that list (because from collections import OrderedDict was meant to be absorbed), so collections.abc matches too and the whole import is swallowed — binding nothing, or None.

Contrast that proves it is the root rule and not module resolution: from xml.etree.ElementTree import Element works and prints its shim note, because xml is not on the consumed list.

The fix is to test the full dotted path, not the root — or to resolve the submodule first and only fall back to the consume rule when nothing resolves.

Why it matters

7 of 48 corpus files stop at unknown base class Mapping (html5lib/_utils.py, _trie/{__init__,_base,py}.py, serializer.py, treebuilders/__init__.py, treewalkers/__init__.py), and every one of them writes from collections.abc import Mapping. The Track B shim lib/rtl/mimic_collections_abc.py is landed and differential-tested against CPython, but it moves zero of those files until this is fixed, because that spelling never reaches it.

RESOLVED. The fix, and the corpus moved.

PyImportRootIsConsumedOnly became PyImportIsConsumedOnly(root, dotted) and the collections rule now matches the full dotted path:

Result := PyImportRootHasNoBackingUnit(root) or
          (CaseEqual(root, 'collections') and CaseEqual(dotted, 'collections'));

The two arms deliberately test different things, and that is the point rather than an inconsistency left behind — it is written into the code so the next reader does not "fix" it. PyImportRootHasNoBackingUnit is keyed on the ROOT because those roots are absent as a whole SUBTREE: there is no typing.io unit either, so every submodule is equally unresolvable and consuming the lot is right. The collections rule is about what ONE MODULE EXPORTS, which says nothing whatever about its submodules, so it has to match the full path.

Checked the sibling first, as [[bug-n-from-sys-import-fails-while-import-sys-works]] warns: PyImportRootPlainIsConsumedOnly deliberately EXCLUDES collections so the plain import spelling already reached the resolver for a submodule. This therefore closes the asymmetry between the two spellings rather than reintroducing one.

Measured on the corpus — all 7 files moved

Same file, same include roots, only the compiler differing (pinned v357 vs a self-hosted fixedpoint at HEAD):

file pinned HEAD
_trie/__init__.py unknown base class Mapping OK
_trie/_base.py unknown base class Mapping OK
_trie/py.py unknown base class Mapping OK
_utils.py unknown base class Mapping undefined variable (__name__)
serializer.py unknown base class Mapping undefined variable (__name__)
treebuilders/__init__.py unknown base class Mapping undefined variable (__name__)
treewalkers/__init__.py unknown base class Mapping undefined variable (__name__)

3 compile clean, 4 advance to a new and unrelated wall. The Mapping row is gone from the ladder. Track B's mimic_collections_abc shim is live — a class D(Mapping) with __getitem__/__iter__/__len__ now gives output byte-identical to CPython, and the build prints note: collections_abc -> mimic_collections_abc (shim, subset).

The 4 that moved all stop at ONE site: _utils.py:126 does baseModule.__name__ after from types import ModuleType, and types does not resolve, so the attribute degrades to a bare name. Filed as [[bug-n-an-attribute-on-an-unresolved-import-degrades-to-a-bare-name]].

Two pre-existing bugs found regression-checking the consumed arm

Both identical on PXX_STABLE, both unaffected by this change (which was the point of checking): from collections import OrderedDict is undefined variable, and from collections import Counter silently answers 0 for every key — [[bug-n-from-collections-import-counter-binds-something-that-always-answers-zero]] (N, p58). The second breaks the consume rule's own promise that an unsupported name walls VISIBLY.

Log