B020: don't flag names bound in nested destructuring patterns (#521)

[aashish254]

B020 matched any name inside the loop target tree against the names loaded by the iterable expression. A name bound inside a nested destructuring pattern is not what the for statement rebinds on each iteration — the iterable expression reads the outer value exactly once, before iteration starts — so safe loops were reported:

for i, (ax, (series, def_name)) in enumerate(zip(axs, series)):
    ...  # B020 wrongly flagged `series`

(The two other shapes from #521: for smoother in smoother.smoothers is a genuine top-level shadow and stays flagged; the sibling-attribute cases were already fixed by #568.)

Fix

check_for_b020 now considers only the top-level bindings of the target (including Starred) as plain-name candidates. Whole dotted paths (for self.a in self.b) keep the path comparison from #568 unchanged, and top-level unpacking shadows like for key, values in values.items() are still reported.

Tests

• New eval cases in tests/eval_files/b020.py: the issue-verbatim two-level unpacking, a one-level unpacking, a starred nested unpacking (all must stay clean), and a top-level tuple binding that must still fire.
• Full suite pytest tests/ green (72 passed, 12 skipped — all skips are _PY version-gated eval files on this local 3.9 runner) including the self-clean flake8 checks; flake8 bugbear.py clean (the eval file intentionally contains flagged loops and is not linted); -k b902 pep8-naming env green; mypy bugbear.py shows no new errors on this env (8-error baseline is Python 3.9-specific — ast.TryStar, CI runs 3.13).
• Regression-scoped witness: B020 output diffed before/after across all 71 eval files — zero flags changed except the nested-unpacking FPs this fix removes.
• # OPTIONS-style hypothesis fuzz tests pass.

Fixes #521

Co-authored-by: Claude noreply@anthropic.com

[cooperlees]

This is muse (muse-spark), posting on behalf of @cooperlees via his coding agent.

Reviewed the diff and verified locally: checked out the PR head and ran the full suite in an isolated venv — 84 passed (plus the 2 pre-existing skips on this 3.14 runner). Also merged current main to resolve the README UNRELEASED conflict (kept all entries) and pushed to b020-nested-unpacking; the merge result is green and the PR is MERGEABLE.

What I checked:

• The #521 false positive is genuinely fixed: for i, (ax, (series, def_name)) in enumerate(zip(axs, series)) is clean on the PR branch, fires on main.
• True positives preserved — I diffed old-vs-new B020 output on 11 target shapes: bare for a in a, top-level tuple shadow (for key, items in list(items)), top-level Starred, subscript targets, same-path attributes (for obj.value in obj.value), and the for smoother in smoother.smoothers case from #521 all still fire; sibling attributes (for self.a in self.b, #568) stay clean.
• The nested-Starred edge (for (*a, b) in a) still fires via the top-level Starred unwrap — no behavior loss there. The only silenced shapes are genuinely nested bindings, which is exactly the intent.
• NameFinder is still used elsewhere (B913UsageFinder subclasses it), so no dead import.
• flake8 self-clean on bugbear.py: clean. mypy: same 4 pre-existing errors as main, no new ones. New eval cases in tests/eval_files/b020.py cover the issue-verbatim shape, one-level and starred nesting, plus a still-firing top-level case.

No changes requested. Approving for merge.