disable flaky LLM judge layer (default-denied every clean wish on llama.cpp); wordlist layers 1-2 remain the gate

app.py

@@ -153,9 +153,15 @@ def _default_wiring(traces_dir: Path) -> dict[str, Any]:
         backend = make_backend("mock")
         backend_name = "mock"
     planner = Planner(backend)
-    # Layer 3 (LLM yes/no judgment) runs through the same backend; layers 1-2 are
-    # the sync precheck the POST handler uses as its fast path.
-    moderator = Moderator(judge=lambda text: judge_fn(text, backend))
+    # Moderation layers 1-2 (charset/length + a thorough wordlist with
+    # confusable/leetspeak/skeleton folding) are the reliable gate and run on
+    # EVERY wish. Layer 3 (an LLM yes/no judge through the same model) is opt-in:
+    # on llama.cpp its grammar-constrained verdict often fails to parse and then
+    # DEFAULT-DENIES every clean wish (June 12: "found a town" rejected as
+    # 'uncertain'). Off by default; GODSEED_JUDGE=on re-enables it.
+    use_judge = os.environ.get("GODSEED_JUDGE", "").strip().lower() in ("1", "on", "true")
+    judge = (lambda text: judge_fn(text, backend)) if use_judge else None
+    moderator = Moderator(judge=judge)
 
     def fast_check(text: str) -> Optional[str]:
         verdict = moderator.precheck(text)