[colbymchenry]

Summary

A Python script's import of a module in a folder above it links again, when that module defines every name the script takes from it. On redis, main lost 137 true edges that 1.6.2 had, and this brings all of them back.

This is a regression from #1921, which hasn't shipped.

Root cause

Confirmed in the code. #1921 made imports resolve from the package roots a project has, instead of any file whose path ends in the module's name. resolvePythonModule tries the roots containing the importer, and an ancestor that is neither a project root, a local root nor the importer's own folder isn't one of them, on purpose.

That loses this layout: redis's modules/vector-sets/tests/*.py do from test import TestCase, generate_random_vector, and modules/vector-sets/test.py is the runner whose folder is first on sys.path. Nothing in the tree says that folder is a root.

Fix

As a last step, resolvePythonModule walks the importer's ancestors for the module, nearest first, passing over packages.

• What proves the module. A name like test or utils may just as well be the stdlib's, so every name the importer takes must be one the file defines at module level. from time import sleep doesn't take a project time.py that has no sleep.
• import helper and from helper import *. They take no name, so they prove nothing and stay unresolved, as on main. No repo below has a candidate of either kind, so this costs nothing measured.
• Roots first. A module a root provides still answers first. Sibling folders are never looked at.
• One answer for every caller. The fallback reads the importer's own from-imports inside the resolver, so context.resolveModuleFile finds the module too, and the outside-the-project check from fix(python): a base imported from outside the project is not a project class of that name #3109 and fix(python): a class-name call doesn't follow a base imported from outside the project #3115 doesn't count it as outside. There's a test for that: a class inheriting from TestCase reaches its method.
• Sync. Such a module binds only while it defines those names, so an edit can link or unlink importers though no file appeared or vanished. An edited module in a folder that is neither a package nor a root, with Python folders below it, now re-opens the edges into it and the failures naming it. Any other edit re-opens nothing.

The [Unreleased] bullet from #1921 gains one sentence. No new bullet, because the regression never shipped.

Validation

Tests: 9 new in python-module-resolution.test.ts, 2 of them on sync, both ways against a fresh index. Without the fallback the 4 that expect a link fail, and the 5 that expect none pass. Here all 9 pass. Taking out each guard (every name, plain imports, packages passed over, the nearest module with the names, the re-open on edit) fails exactly its own test.

A/B against main, kernel on:

Repo	Edges	Added	Removed	Moved	Upgraded
redis @ 7f44037	77,608 → 77,781	173	0	0	44
requests, flask, litestar, mypy, sqlalchemy, polar server/polar, django django/, scikit-learn, matplotlib, and 4 small apps	653,583	0	0	0	0
express (control)	2,296	0	0	0	0
gin (control)	8,073	0	0	0	0

Both controls have byte-identical full dumps.

Every class of change, all in redis:

Edges	Class	Read as
69	imports of TestCase, generate_random_vector and fill_redis_with_vectors	The lost imports, right
68	calls of generate_random_vector and fill_redis_with_vectors	The lost calls, right
36	imports of the file test.py	Right, and new against 1.6.2: a from-import from the same folder gets this edge too
44	extends TestCase, exact-match → import	Same target, now proven by the import

Checked against the regression hunt's list: 137 of 137 are back.

Nothing wrong appears. A trace of the fallback logged every candidate it could see, on all the repos above and on whole django, polar's server, ansible and cpython, whose trees hold 66 files named test.py, utils.py, common.py or helpers.py. It fired only for redis's test.py, from its 36 test files, with every name proven and no candidate rejected.

Cost: redis, median of 3 interleaved runs: 12.0 s on main, 10.7 s here. That gap is noise, so nothing measurable.

Suite: 849 files (843 passed, 6 skipped), 10,389 tests passed, 2 expected failures, 0 failed, 87 skipped, on this exact commit, kernel on, 4 workers, Vitest 4.

Not in this PR

• A package above the script that re-exports the name from its __init__.py isn't proven: only the file's own module-level defs count. Following re-exports here would recurse through the resolver. No repo above has the case.
• A dotted module above the script, from helpers.db import connect, is found by the same walk, but an edit to it doesn't re-open its importers; only a module directly in such a folder does. No repo above has this case either.

🤖 Generated with Claude Code