release the world-decode cache before the block loop

app.py

@@ -265,8 +265,14 @@ def run(world: str, steer_to: str, steer_at: float, seconds: float, seed: int):
 
     t0 = time.perf_counter()
     engine.start(world=WORLDS[world], seed=int(seed))
-    _probe("after-world")  # world decoded and the K/V cache allocated: the state the
-    #                        first FFN linear fails in
+    # Opening a world decodes all 81 of its pixel frames in one call, which is by far
+    # the largest allocation in a run and leaves the caching allocator holding blocks
+    # the block loop cannot reuse (reserved 30.5 G against 13.5 G allocated). Growing
+    # past that OOMs, and a ZeroGPU worker cannot even format the OOM message --
+    # PyTorch asks NVML which processes hold memory, NVML is not available there, and
+    # the run dies on an internal assert instead. Hand the cache back first.
+    torch.cuda.empty_cache()
+    _probe("after-world")
     frames, pending_steer, last_push = [], steer_idx is not None, 0.0
 
     # The worker runs in a thread and reports failure by setting status.state rather