Streaming-zip: default use_plugin_default=True

tools/hf_space/runner.py

@@ -408,16 +408,17 @@ def _validate_zip_streaming(*, api: HfApi, dataset: str, token: str | None,
     val_ver = _validator_version()
     found_sha = _foundation_sha()
 
-    # Deterministic recovery state. When the default CLI loader emits
-    # "FATAL: profile X not registered" on the first zip (which is the
-    # signature of the foundation specs referencing requirement codes
-    # the simready-validate package doesn't have registered), we retry
-    # that zip with --use-plugin. If --use-plugin works, we promote it
-    # to the default for every subsequent zip in this run so we don't
-    # pay the retry cost N times. This is the same strategy the agent
-    # would have chosen, just done in code so we don't burn an LLM
-    # call per zip on a deterministically-known failure mode.
-    use_plugin_default = False
+    # Always use --use-plugin in the streaming path. With the patched
+    # foundation wheel installed, SimReadyPlugin's on_startup() loads
+    # profiles via simready.validate.impl.loader at omni.asset_validator
+    # import time — fully populated ProfileRegistry before validate.py
+    # main() runs. The default CLI-loader path still attempts a
+    # second `load_validation_implementation` call which races with
+    # the plugin's registration and ends up failing in subtle ways
+    # (we've observed FATAL: profile X not registered even though
+    # the plugin succeeded — different code path, different bug).
+    # Skip the doomed-first-attempt entirely.
+    use_plugin_default = True
     issue_filed_for_registration_bug = False
     # Once an issue-filing attempt 404s (token can't reach the repo,
     # repo wrong, permissions missing), don't try again this run —