test(hygiene): stop the suite writing to the operator's real data dir

[Cipher208]

Several modules resolve their data directory once, at import time, from MCP_MEMORY_DATA_DIR: shared/saga/impl/storage.py (SAGA_DIR), shared/read_only.py (which builds the module-level read_only_replica singleton on import) and features/auth/__init__.py. Importing any of them while the real value was in place was enough to write to the live directory — no test body had to do anything wrong, and hermetic_global_db could not help, because by the time a fixture runs those constants are already resolved.

What the suite was leaving behind

Measured on ~/.mcp-ariel-memory before this change:

path	left behind
replica/memory.db	14 MB, rewritten on every run by the lifespan
sagas/	15875 state files, 63 MB
<repo>/data/auth/keys.enc	704 KB, +~860 B per run

The fix

tests/conftest.py now publishes a session temp dir as MCP_MEMORY_DATA_DIR before pytest imports the first test module, and the session fixture removes it and restores the operator's value on teardown. hermetic_global_db reuses that directory instead of making its own.

After the change, a full run leaves all three paths byte-identical: same mtime, same size, same file count.

Two tests were relying on the variable being unset

• test_backup_saga_success patched pathlib.Path.home to fake a data dir. That only worked while MCP_MEMORY_DATA_DIR was unset and the fallback was reachable; the variable wins. It now names its own dir through the variable the saga actually reads.
• test_backup_saga_compensation passed vacuously: with no source database the first step reported skipped_no_source, so backup_path was absent and the assertion that the backup was removed never ran. It now requires the backup to have existed — and finds one.

test_auth_backup.py also constructed APIKeyAuth() / BearerAuth() bare, falling back to the CWD-relative data/auth/*.enc, which no environment variable can redirect — hence the separate fix there. Each test passes its own tmp_path, and test_api_key_list asserts == 2 rather than >= 2: it was being satisfied by the accumulated history, not by itself.

Tests

Full suite: 2087 passed.

Not addressed here, reported instead

• The same scratch dir holds 1.6 GB of tmp* directories left by tests that call tempfile.mkdtemp() without cleaning up. Different mechanism, own PR.
• data/auth/keys.enc from earlier runs is still on disk. It is gitignored (*.enc), so nothing was ever committed.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Repository: Cipher208/a-memory/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: c3bca5f4-6d48-4396-a8ec-9276c80c148a

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}