fix(tests): order-dependent parser-test isolation — the bare-name import-state leak (#627)

[gadievron]

What

pytest tests/parsers failed 12 tests on a clean master (4 in test_issue309_py_callgraph_constructs.py, 8 in test_issue440_mro_import_bases.py) while every file alone, the two together, and the full suite passed — the exact receipt from the issue. Closes #627.

Root cause (bisected)

The poisoner is tests/parsers/c/test_empty_seed_keep_all.py. It execs parsers/c/test_pipeline.py, which — like parsers/python/parse_repository.py and the PHP and Ruby pipeline scripts — imports its parser machinery under bare module names (function_extractor, call_graph_builder, repository_scanner, unit_generator) resolved via a sys.path entry. The exec leaves the C versions in sys.modules; a later parse_repository(language="python") then gets the C FunctionExtractor (first-import wins), whose statistics lack standalone_functions — the KeyError behind all 12 failures. The full suite passes only because some earlier test imports the python bare names first: order decides the winner, which is exactly the flakiness class.

The bisection: directory halving with a failing file as the detector → tests/parsers/c → single-file scan → test_empty_seed_keep_all.py; the pair (poisoner + both failing files) reproduces 12/12 deterministically.

The fix (isolation, not a reorder)

Snapshot sys.path and sys.modules at loader ENTRY and restore around the exec — new keys removed, replaced keys restored. Applied to the loaders that exec bare-importing pipelines: c, zig, php, ruby (Go, JavaScript, and Swift use qualified imports and do not poison the bare-name cache; their loaders still leak sys.path state — a broader sweep is noted as follow-up).

Testing

• The regression test (test_issue627_import_state_isolation.py) reproduces the poisoning within one test — no invocation order needed: exec the C loader, then run a python parse_repository resolution; it asserts the pkg-defined re-export resolves and that the loader leaves no import state.
• Stash-RED verified: with the fix reverted the regression test fails (2 failed), restored it passes.
• The receipt inverts: pytest tests/parsers → 918 passed, 1 skipped, 0 failed (was 12 failed, 805 passed).
• Full suite: 4337 passed, 34 skipped, 2 failed — both test_llm_sdk_contract_floor pins, a local-environment artifact (installed anthropic 1.3.0 vs the requirements.txt pin 1.7.0; CI installs the pin), not master failures. Corrected 2026-09-21.

Coordination

• feat(llm): the per-phase request-side thinking policy — exposed, gated, recorded (#625) #658/feat(llm): request-side thinking control, effective-policy records, dropped-block diagnostic (#625) #660 (thinking control): no shared files; the loaders touched here are parser-test-only.
• The sibling set (c, zig, php, ruby loaders) is derived from each pipeline's import style — Go/JavaScript/Swift use qualified imports and are noted as a follow-up sweep for path-state-only leaks.