Package / sibling-module resolution is now the largest corpus wall

Measured by Track B on v345, 2026-08-17, over the 48 non-test files of the NilPy corpus ladder — first error per file:

webencodings + constants + _utils are 11 files between them, and they are not missing shims — they are a package importing its own siblings. That is resolution work in parser.inc, Track A, and it is now the biggest single lever on the ladder.

Why this is the lever and six was not, despite six gating more files

six gated 15 files and landing mimic_six moved the compile count not at all — 4/48 before and after. That is not a failure and it was predicted: a file stops at its first missing import, so clearing one wall exposes the next.

The number that moved: 13 files had six as their first wall; now 0 do. six does not appear in the table above.

Reporting rule that follows, and it applies to the whole campaign: with stacked walls, compile count is a LAGGING indicator and walls-cleared is the LEADING one. Judging a fix by files-compiling will call a real unblock a zero, and will keep doing so until the last wall on some file falls. Report both, lead with walls.

Related, already filed

• bug-n-a-subpackage-directory-does-not-resolve-as-a-module — unblocked by bug-a-a-python-module-s-identity-is-its-name-not-its-file (030ce07ea), waiting on Track N staffing for its .npy half. Likely the same ground.
• bug-n-a-class-base-that-is-an-expression-does-not-compile (N, p45) — the only wall on six.with_metaclass; smaller than it looks, since html5lib's getMetaclass returns plain type unless a debug flag is set, so the real path asks for no metaclass at all. It needs base-expression evaluation, not metaclass support.

2026-08-17 (frank2, Track A) — RESOLVED. The cause is the DOT LEVEL, and two of the three named walls were not what the table said.

Re-derived from the corpus rather than inherited, and the ticket's framing — "a package importing its own siblings" — turned out to name three different situations, only one of which was a defect.

constants and _utils: the real bug is .., not siblings

from .constants import X already worked. What failed is from ..constants import X — a PARENT-relative import, which is what html5lib's subpackages actually write:

html5lib/filters/lint.py:6      from ..constants import namespaces, voidElements
html5lib/treewalkers/__init__.py from ..constants import ...

PyRelativeImportLevel already returns the dot COUNT, and every caller discarded it — the note on that function says the dots are "simply skipped" and calls level 1 "the closest available one for ..". It is not close enough: level ≥ 2 was then resolved against the SUBpackage's own directory, where the module does not exist, while the absolute from html5lib.constants import ... on the same tree resolved fine.

Minimal repro against the CPython oracle:

pkg/constants.py      TOP = 7
pkg/sub/leaf.py       from ..constants import TOP
  pxx:     error: no unit named constants
  CPython: 7

Fix: publish the level (PyRelativeLevel, claimed-and-cleared in ParseUsesUnitBody exactly like PyDottedImport) and resolve the sibling probe against PyRelativeDir(CurUnitDir, level) — level 0/1 mean "here", each dot beyond the first climbs one package, and a climb above the root returns '' so the caller simply finds nothing rather than resolving against something arbitrary.

webencodings: NOT a compiler bug — a scan artifact

The table's largest entry (6-7 files) is cross-package, not sibling: tinycss2/bytes.py does from webencodings import ..., and the ladder scan passes only the file's OWN candidate root on -Fu. With both roots present it resolves and the wall moves to undefined variable (CodecInfo) — the known mimic_codecs surface gap, already Track B's.

So that row measured the harness, not the compiler. Worth fixing in the scan method rather than in parser.inc: it is the equivalent of leaving a library off PYTHONPATH. Not changed here — the scan is Track B's instrument and re-defining their measurement mid-campaign is not mine to do; flagged instead.

Result, led by walls-cleared per the ticket's own rule

Measured over the 15 corpus files that actually contain a parent-relative import — the population this fix can affect — A/B'd against a baseline built from HEAD minus this diff (not pinned):

5 walls cleared, 0 regressions, compile count unchanged at 0 for this population. constants and _utils disappear from its first-wall table entirely. Reported walls-first exactly because the compile count is flat and would read as a zero.

Test

test/test_nilpy_relative_parent_import.npy + package test/relpkg/, enumerated in test-nilpy. It imports BOTH spellings from one file — from ..constants and from .peer — so it fails whichever direction the level is ignored in, rather than only catching a missing climb. Output matches CPython (42), and it is confirmed RED on the HEAD-minus-diff baseline.

Note on the frontend edit

compiler/pyparser.inc is Track N's file and N is unstaffed; the edit there is four lines that publish the already-computed level to the resolver, with the resolution logic itself in parser.inc (Track A). Flagged rather than assumed.

Gate

make compiler/pascal26 + repro + tools/gate.sh quick. A .npy package importing a sibling module and a subpackage, both spellings, both import orders — the order matters, see 030ce07ea, where whichever spelling lost the race was the one that broke.

Log

• 2026-08-17 — resolved, commit 546c03e02.

CORRECTION — the webencodings row measured the HARNESS, not the compiler

Found by Track A while fixing this (2026-08-17). The coordinator filed this ticket citing that row as the largest lever; that was wrong, and the error was in the instrument rather than in anyone's reading of it.

webencodings is cross-package, not sibling: tinycss2/bytes.py does from webencodings import ..., and the ladder scan passes only the file's OWN candidate root on -Fu. Supply both roots and it resolves immediately — the wall moves to undefined variable (CodecInfo), the known mimic_codecs gap already tracked on Track B.

So 6-7 files of the "biggest wall" were the equivalent of leaving a library off PYTHONPATH. Track A did NOT change the scan — it is Track B's instrument and redefining another lane's measurement mid-campaign is not Track A's call — so this is routed rather than fixed.

Why it matters beyond one row: while that row stands, it will keep ranking as the top lever and keep sending Track A after a resolution bug that does not exist. A measurement artefact that survives is worse than a missing measurement, because it is actionable and wrong.

What the fix actually was: the DOT LEVEL

from .constants import X already worked. The defect was from ..constants import X — the parent-relative form html5lib's subpackages write. PyRelativeImportLevel already returned the dot count and every caller discarded it; its own note said the dots are "simply skipped" and called level 1 "the closest available one for ..". Level >= 2 then resolved against the subpackage's directory, where the module is not, while the absolute spelling on the same tree resolved fine. Fixed by publishing the level and climbing level-1 packages.

Result, walls-first

15 corpus files carry a parent-relative import; A/B'd against HEAD-minus-the-diff (not pinned — see the roster's standing note on why pinned is not a baseline):

5 walls cleared, 0 regressions, compile count flat at 0. constants and _utils leave the first-wall table entirely. Read by compile count this is a zero, which is exactly what the lagging/leading rule above predicts.

Remaining walls in this population are undefined variable (digits) (i.e. string.digits) and xml_etree_elementtree — library/shim work, not Track A.

2026-08-17 — the wall table quoted above is SUPERSEDED. It came from a scan that passed only the scanned file's own -Fu root, so cross-package imports (tinycss2 -> webencodings) recorded as compiler walls. Corrected table and the tool that now produces it (tools/nilpy_ladder.py): [[bug-t-the-ladder-scan-passes-only-one-root-so-cross-package-imports-read-as-walls]]. The webencodings and constants rows were artefacts and are gone; the real top two are undefined variable (digits) (8 files) and undefined variable (CodecInfo) (7 files).