correct the empty_cache comment: autograd was the OOM

app.py

@@ -262,12 +262,11 @@ 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))
-    # 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.
+    # Opening a world decodes all 81 of its pixel frames in one call, the largest
+    # allocation in a run, and leaves the allocator holding ~17 GB the block loop has
+    # no use for (reserved 30.5 G against 13.5 G allocated). Handing it back is not
+    # what fixed the OOM this Space started with -- that was autograd, see
+    # `harden_worker` -- but there is no reason to sit on it either.
     torch.cuda.empty_cache()
     torch.cuda.reset_peak_memory_stats()
     frames, pending_steer, last_push = [], steer_idx is not None, 0.0