fix(witan-code): stop _git_context leaking across test boundaries

[blarghmatey]

What & why

tests/test_crossrepo.py::test_repo_scope_route_and_fanout failed in the full local suite but passed alone and in CI (tk-witan-code-test-repo-scope-route-and-fanout-fail-348361).

The task filed against this suspected residue in a shared store. It is not that. The cause is server._git_context (witan_code/server.py:107), which memoizes detect()/store_branch() for a 2 second TTL that outlives a single test.

The failure sequence:

1. test_bridge.py sets WITAN_REPO=https://github.com/test/repo-a and memoizes it.
2. If test_crossrepo starts inside that 2s window, its own monkeypatch.setenv("WITAN_REPO", "") — whose entire purpose is to make detect() return None so the query fans out — is silently ignored.
3. _resolve_clients(None) therefore takes the single-repo branch instead of the fan-out branch. repo-a's store does exist in the test's tmp code dir, so it returns one client.
4. The fan-out collapses to repo-a alone, and repo-b goes missing — exactly the reported Extra items in the left set: 'https://github.com/test/repo-b'.

Whether the leak lands is a race against wall-clock, decided by how fast the preceding tests happen to run. That is why it reproduced locally and not in CI — and why a green run does not prove absence.

The fix

Clear _git_context between tests in an autouse fixture, directly alongside the identical _fresh_identity fixture that already exists for identity._cache.

This was already a known hazard: 13 hand-rolled srv._git_context.clear() calls are scattered across the suite working around it per-test. test_crossrepo was simply one of the tests that did not know it had to. The clears that run mid-test (after a checkout or an actor switch) are still load-bearing and are left in place.

No production behaviour changes. The TTL is unchanged and still amortizes git subprocesses across an agent turn. The source comment claimed the TTL was "short enough that no test needs to know about it" — the assumption that made this trap invisible. It now says the opposite.

Verification

Both directions were mutation-checked by toggling autouse off and re-running:

	with fixture	without fixture
new regression pair	passes	fails — a previous test's git context leaked
seeded test_crossrepo leak scenario	passes	fails — reproduces the original Extra items in the left set: 'repo-b' verbatim

The regression test is a deliberate ordered pair: the first leaves residue, the second asserts it is gone. Delete the fixture and the second fails — so the fixture cannot be quietly removed later.

Full suite: 438 passed (436 + the 2 new tests). prek clean on all changed files.

[Copilot]

Pull request overview

This PR fixes a flaky witan-code test failure caused by the module-level server._git_context TTL cache persisting across test boundaries, allowing earlier tests’ memoized repo/branch detection to affect later tests.

Changes:

• Update witan_code.server’s cache commentary to explicitly document cross-test leakage risk and the expected test mitigation.
• Add an autouse pytest fixture to clear server._git_context before/after each test (mirroring the existing identity cache reset pattern).
• Add a small regression pair in test_branches.py to ensure git-context memoization does not persist into subsequent tests.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
mcp/servers/witan-code/witan_code/server.py	Clarifies the intent and testing implications of the _git_context TTL cache.
mcp/servers/witan-code/tests/conftest.py	Adds an autouse fixture to clear _git_context between tests to prevent leakage.
mcp/servers/witan-code/tests/test_branches.py	Adds regression tests asserting _git_context is cleared between tests.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.