fix(test): use relative cache expiry to avoid CI timing fluke

On freshly-started CI runners time.monotonic() can be < 300 s, so
setting _models_cache_ts = 0.0 left the cache appearing valid and the
second pipes() call returned early without clearing _icons_synced.
Subtract the full TTL + 1 from the stored timestamp instead so the
cache is guaranteed expired regardless of system uptime.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: sena-labs

test_pipe.py

@@ -2217,8 +2217,10 @@ def _ttl_reg_get(url, *args, **kwargs):
 _pipe_sync_clear._icons_synced.add("openai/gpt-4o")
 _assert(len(_pipe_sync_clear._icons_synced) == 1, "_icons_synced: populated before cache expire")
 
-# Expire cache and call pipes() again — _icons_synced must be cleared
-_pipe_sync_clear._models_cache_ts = 0.0
+# Expire cache and call pipes() again — _icons_synced must be cleared.
+# Subtract more than the TTL from the stored timestamp so the cache is
+# expired regardless of how small time.monotonic() is on a fresh CI runner.
+_pipe_sync_clear._models_cache_ts -= mod._MODELS_CACHE_TTL + 1
 with patch.object(_pipe_sync_clear._session, "get", return_value=_mock_models_resp_sc):
     _pipe_sync_clear.pipes()