fix: clear compat collector per session and set LITELLM_MASTER_KEY for cron proxy

- Add pytest_sessionstart hook in tests/claude_code/conftest.py that
  clears the module-level _COLLECTOR so results don't leak across
  pytest.main() invocations within the same process.
- Export LITELLM_MASTER_KEY=${PROXY_API_KEY} when launching the cron
  VM proxy so the auth token the tests send actually matches what
  the proxy expects.

tests/claude_code/conftest.py

@@ -146,6 +146,17 @@ def pytest_runtest_makereport(item, call):
     )
 
 
+def pytest_sessionstart(session):
+    """Clear collected results at the start of each session.
+
+    The `_COLLECTOR` is a module-level singleton, so it survives across
+    `pytest.main()` invocations within the same Python process. Without
+    this reset, results from a prior session would leak into the next
+    run's `compat-results.json` artifact.
+    """
+    _COLLECTOR.items.clear()
+
+
 def pytest_sessionfinish(session, exitstatus):
     """Write the structured results artifact at end of session."""
     artifact_path = os.environ.get(RESULTS_ARTIFACT_ENV) or DEFAULT_ARTIFACT_PATH

tests/claude_code/cron_vm/run_daily.sh

@@ -204,7 +204,7 @@ log "starting proxy on :${PROXY_PORT}"
 # SIGTERM the whole tree by passing the pgid as a negative pid. We
 # write that pid to a file so cleanup() doesn't need to remember a
 # variable that might be stale by the time the trap fires.
-setsid bash -c '
+setsid env LITELLM_MASTER_KEY="${PROXY_API_KEY}" bash -c '
   echo "$$" > "$0"
   cd "$1"
   exec "$2" run litellm --config "$3" --port "$4"