01 — Isolate synth state in tests

Priority: P0
Depends on: none
Size: S (30–60 min)
Files: tests/test_synth_pipeline.py
PR title: fix: isolate synth state in tests

Read first

tests/test_synth_pipeline.py (the whole file, ~450 lines)
sessionwiki/synth/pipeline.py, functions _resolve_state_file, _load_state, _save_state (around lines 115–180)

Problem

tests/test_synth_pipeline.py calls synthesize_new_sessions() without passing state_file. The function falls back to the real repository state file:

STATE_FILE = REPO_ROOT / ".llmwiki-synth-state.json"

Running the test file overwrites that file with six fixture keys:

proj/2026-03-01-flickering-orbiting-fern.md
proj/2026-04-01-flickering-orbiting-fern.md
proj/2026-04-01-slug-0.md
...

The real file has 800+ keys. After a test run, the next sessionwiki synthesize re-processes the whole corpus. That is ~45 minutes and about $55 by the current estimate.

Reproduced with:

cp .llmwiki-synth-state.json /tmp/state.before.json
.venv/bin/python -m pytest tests/test_synth_pipeline.py -q
python3 -c "import json; print(len(json.load(open('.llmwiki-synth-state.json'))))"

The last command prints 6 instead of 800+.

Required change

Add one autouse fixture at the top of tests/test_synth_pipeline.py, after the imports:

@pytest.fixture(autouse=True)
def _isolated_synth_state(tmp_path, monkeypatch):
    """Keep the real .llmwiki-synth-state.json out of test runs."""
    monkeypatch.setattr(
        "sessionwiki.synth.pipeline.STATE_FILE",
        tmp_path / "synth-state.json",
    )

_resolve_state_file() reads the module-level STATE_FILE at call time, so monkeypatching the attribute is enough.

Add one regression test in the same file:

def test_synth_state_is_isolated(tmp_path):
    from sessionwiki.synth import pipeline

    assert pipeline.STATE_FILE.parent == tmp_path

Do not change sessionwiki/synth/pipeline.py. The production fallback is correct; the tests must isolate themselves.

Verification

cp .llmwiki-synth-state.json /tmp/state.before.json
.venv/bin/python -m pytest tests/test_synth_pipeline.py -q
cmp .llmwiki-synth-state.json /tmp/state.before.json && echo "state unchanged"

cmp must print nothing before state unchanged.

Then run the full suite and repeat the comparison:

cp .llmwiki-synth-state.json /tmp/state.before-full.json
.venv/bin/python -m pytest tests/ -q
cmp .llmwiki-synth-state.json /tmp/state.before-full.json && echo "state unchanged"

If the full suite still changes the file, another test file writes to the default state. Find it with:

grep -rn "synthesize_new_sessions(" tests | grep -v "test_synth_pipeline"

Add the same fixture to that file. Do not weaken this card to only the first file.

Acceptance criteria

[ ] The autouse fixture exists in tests/test_synth_pipeline.py.
[ ] test_synth_state_is_isolated passes.
[ ] tests/test_synth_pipeline.py leaves .llmwiki-synth-state.json byte-identical.
[ ] The full tests/ suite leaves .llmwiki-synth-state.json byte-identical.
[ ] pytest tests/ -q exits with code 0.
[ ] CHANGELOG.md has one line under ## [Unreleased].

Out of scope

Refactoring _load_state or _save_state.
Changing the production state path.
Fixing other test files unless the full-suite check proves they touch the state file.

Rollback

Revert the fixture and the new test. No production code changes.