Problem

Four tests fail intermittently under pytest-xdist whenever the
openai-codex token-refresh test happens to run on the same worker
before them:

FAILED tests/hermes_cli/test_auth_commands.py::test_auth_remove_accepts_label_target
FAILED tests/hermes_cli/test_auth_commands.py::test_auth_remove_prefers_exact_numeric_label_over_index
FAILED tests/hermes_cli/test_opencode_go_in_model_list.py::test_opencode_go_appears_when_api_key_set
FAILED tests/hermes_cli/test_overlay_slug_resolution.py::test_kimi_for_coding_overlay_uses_hermes_slug

Each one observes an extra phantom openai-codex credential it never
seeded. For example:

>   assert len(entries) == 1
E   AssertionError: assert 2 == 1

Root cause

Production _refresh_codex_auth_tokens writes refreshed OAuth tokens
back to $CODEX_HOME/auth.json (defaulting to Path.home() / ".codex" / "auth.json") so the Codex CLI / VS Code extension stay in
sync with Hermes:

# hermes_cli/auth.py
def _refresh_codex_auth_tokens(tokens, timeout_seconds):
    refreshed = refresh_codex_oauth_pure(...)
    _save_codex_tokens(updated_tokens)
    # Write back to ~/.codex/auth.json so Codex CLI / VS Code stay in sync.
    _write_codex_cli_tokens(refreshed["access_token"], refreshed["refresh_token"], ...)

tests/agent/test_credential_pool.py::test_try_refresh_current_updates_only_current_entry
mocks the inner refresh_codex_oauth_pure but not the outer
_refresh_codex_auth_tokens, so the writeback runs for real. The test
only sets HERMES_HOME, never CODEX_HOME, so
_write_codex_cli_tokens falls back to Path.home() / ".codex" / "auth.json".

Under the project's standard HOME=$(mktemp -d) test wrapper that HOME
directory is shared by every test in the run. The leaked
auth.json therefore persists for the rest of the session, and any
later test that calls load_pool("openai-codex") re-imports those
stale tokens via _seed_from_singletons →
_import_codex_cli_tokens, ending up with the phantom credential.

Fix

Pin CODEX_HOME to a per-test scratch dir from the autouse
_isolate_hermes_home fixture, mirroring the existing HERMES_HOME
treatment:

fake_codex = tmp_path / "codex_test"
fake_codex.mkdir()
monkeypatch.setenv("CODEX_HOME", str(fake_codex))

The writeback now lands inside tmp_path, which pytest cleans up
between tests, so no later test ever sees the file. Add a regression
test in tests/agent/test_credential_pool.py that exercises
pool.try_refresh_current() for openai-codex and asserts
Path.home() / ".codex" / "auth.json" is not created.

Tests that need a specific CODEX_HOME continue to override it via
monkeypatch.setenv(...) as before.

Verification

$ python -m pytest tests/ -q
...
11270 passed, 79 skipped in 187.13s

(All four originally failing tests pass under both serial and xdist
runs; the new regression test passes; nothing in the rest of the
suite regresses.)