[DeusData]

What

#2704 publishes a Python package's re-exports from the module's ordered namespace facts. An unknown statement in those facts withdraws every re-export recorded before it. Two kinds of statement were recorded as unknown although they cannot rebind a name:

"""Module docstring."""                  # a string-literal statement
from __future__ import annotations       # recorded as UNSUPPORTED
from .elements import nulls_first
nullsfirst = nulls_first
"""Synonym for the :func:`.nulls_first` function."""   # an attribute docstring: EFFECT

lib/sqlalchemy/sql/expression.py lost all 149 of its re-exports to the attribute docstrings after them. Only plain one-line strings were proven harmless.

This change:

• String-literal statements are no effect (extract_imports.c). This covers any quoting (single, double, triple), the r / b / u / rb prefixes, implicit concatenation, and an f-string without a replacement field, as a module, class or attribute docstring. An f-string with a {...} field stays an effect, alone or inside a concatenation, because its __format__ / __str__ may run code. The same predicate covers string-literal values, so __all__ = ['escaped\n'] is no longer an effect either. Its __all__ stays unknown, because an escaped name is not a literal export name.
• from __future__ import x [as y] binds like a from-import of module __future__. Tree-sitter parses it as its own node type, future_import_statement, without a module field, which is why it was UNSUPPORTED. from __future__ import * is a syntax error in Python and in the grammar, so the file stays a parse unknown.
• CBM_PY_NAMESPACE_FACTS_VERSION 1 → 2. The version travels in the lsp_surface "py" payload. A stored version-1 payload no longer decodes, so the first incremental run on an existing index declines the closure repair (python_namespace_metadata_invalid). It goes through the existing FORCE_FULL_REINDEX path once: one full re-index, after which the facts are current. Only Python payloads are affected, and SURFACE_CODEC_VERSION is unchanged. The wire shape is unchanged; version 2 changes which statements extraction proves harmless.

Other languages are unchanged.

Known limit: sql.cast, sql.select, sql.and_

lib/sqlalchemy/sql/__init__.py (and sqlalchemy/__init__.py, sqlalchemy/orm/__init__.py) still end with a top-level __go(locals()) call. That call is an unknown effect and withdraws those packages' re-exports. Those three names therefore still resolve by the old fallback. Proving that call harmless is a separate change.

Evidence

Clean builds (build dir removed before make), par, repeat 2, noise 0 on every repo. Base = #2704's head 67e53d0; head = this branch bf56fcf.

repo	moved
sqlalchemy	CALLS +6 / -2, 12 same-target strategy upgrades; nothing else
sentry, ansible, django	byte-identical

Judged, all 20 moved edges (snippets):

• Added CALLS (5 of 6 correct):
  • expression.select(...) ×2 → sql._selectable_constructors.select (lsp_module_attr); each replaces the guess WriteOnlyCollection.select (suffix_match). Both removals are correct.
  • expression.Join.__init__(...) → selectable.Join.__init__.
  • expression.ClauseList(...)._annotate(...)._set_propagate_attrs(...) → SupportsWrappingAnnotations._annotate, ClauseElement._set_propagate_attrs (MRO).
  • Not correct: class array(expression.ExpressionClauseList[_T]) → ColumnOperators.__getitem__ (lsp_operator_dunder). A subscript of a class runs __class_getitem__, not the instance __getitem__. This is an existing operator-dunder shape of the Python LSP, reached now because expression.ExpressionClauseList resolves. It is a separate fix.
• Upgrades (12, all the same target): unique_name / suffix_match (0.38-0.75) → lsp_module_attr / lsp_constructor (0.88-0.92), all through expression.<X>. The targets are the savepoint clauses, BindParameter, Grouping, BinaryExpression, ClauseList, _select_iterables, text, type_coerce and null.

Namespace census (the facts as extraction stores them, replayed with #2704's fold):

• This change frees exactly one namespace with consumers in these four repos: sqlalchemy sql/expression.py, with 149 re-exports. 141 of those names are imported from it elsewhere in the repo (269 importer/name pairs).
• In sentry, six test and helper modules gain 19 names; in ansible, executor/stats.py gains 2. Nothing imports those names from those modules.

Tests

• extraction: python_namespace_facts_string_statements_and_future_imports: every string form after an import, the class and attribute docstrings, and three __future__ shapes (plain, as, parenthesized). Controls: an f-string field, a concatenated f-string field and a call stay effects; from __future__ import * is a parse unknown.
• pipeline: pipeline_python_package_reexports_resolve_by_lsp adds two cases, each over both drivers × full and incremental: a re-export between docstrings and a __future__ import resolves, and one before an f-string with a field does not.
• pipeline: pipeline_python_namespace_codec_rejects_non_authoritative_rows: a version-1 payload is rejected.
• extraction: python_namespace_facts_multiple_bindings_and_unsupported_effects now pins 13 facts instead of 14. The effect fact for __all__ = ['escaped\n'] existed only under the old string rule, and the __all__ unknown that follows it is unchanged.
• RED on revert: with the two production files back at the base, all three fail. Each of the eight string forms is its own effect fact, the two __future__ statements are unsupported, the version is 1, and the pipeline re-export does not resolve. GREEN with the change.
• The 31 affected suites pass on the branch merged with main d4980ad: 4,119 tests, 0 failures (clean ASan/UBSan test-runner).
• lint-memory-core reports no growth. cppcheck 2.20 (the CI version) and clang-format are clean.