fix: preserve tested_by edges for Python test files under tests/

[sihaoqiaouzi]

Summary

is_test_path() treats tests/ as a test-source root for Swift, Rust, and PHP — but Python is missing from _TEST_DIRECTORY_EXTENSIONS. Python files under a tests/ directory are only recognized when the basename carries a test_ prefix or a _test suffix.

That leaves a common layout broken: a project whose suite lives in tests/ under names like run_tests.py or stage4_e2e.py gets every production-to-test tested_by edge dropped at merge time, because the pair is read as production-vs-production and discarded as semantically broken. The merge reports those edges as dropped, and the resulting graph shows zero test coverage for a repo that has a full test suite.

Reproduction

A 30-file Python project — tests/run_tests.py, tests/stage3_tests.py, tests/stage4_e2e.py, production under data_agent/:

• The file-analyzers classified all three correctly and emitted 25 correct tested_by pairs.
• merge-batch-graphs.py classified all three as production files and dropped every pair.
• Resulting graph: 0 tested_by edges.

With this patch, on the same project:

	before	after
is_test_path on the 4 real tests/*.py files	4/4 wrong	12/12 correct
tested_by edges in graph	0	17
production files tagged tested	0	12
plugin unittest suite	91 passed	93 passed

Change

-_TEST_DIRECTORY_EXTENSIONS: frozenset[str] = frozenset({".swift", ".rs", ".php"})
+_TEST_DIRECTORY_EXTENSIONS: frozenset[str] = frozenset({".swift", ".rs", ".php", ".py"})

Plus the matching docstring/comment update and two new tests:

• test_python_tests_directory — Python naming-convention false-negatives (run_tests.py, stage3_tests.py, stage4_e2e.py, gen_data.py, helpers.py).
• test_python_non_test_paths — guards against a tests substring appearing elsewhere in the path (data_agent/core/sqlite_utils.py, mypkg/testing/helpers.py, src/latest.py).

The Swift/Rust/PHP cases added in #689 still pass.

Relation to the other tested_by reports

Both open issues below are about dropped tested_by edges, but each has a different root cause. This patch is not a duplicate of either.

• tested_by path-convention linker produces 0 edges for flat tests/ dir against a nested source package #586 — a flat tests/test_*.py layout against a nested src/<pkg>/<subpkg>/ package produces 0 edges. There the path does match the convention, but the generated candidates never reach the real source file. This patch does not touch that code path.
• merge-batch-graphs.py drops tested_by edges to test files that don't match naming conventions — trust the analyzers' 'test' tag #595 — a verification suite outside any tests/ root (scripts/verify-*.mjs) is never classified as a test. That case requires trusting the analyzers' test tag. This patch keeps the existing path-convention rule and only completes its extension list, so it stays correct — and complementary — if the tag-based approach lands later.

Testing

cd <repo>
python -m unittest tests.skill.understand.test_merge_batch_graphs
# Ran 93 tests ... OK