Bound graph predictor grid to runtime

runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py

@@ -454,6 +454,13 @@ def main() -> None:
         config_knobs["no_cache"] = args.fresh_predictor_cache
         if args.align_real_graph_runtime:
             config_knobs["num_blocks"] = REAL_NUM_BLOCKS_BY_CONFIG[(config.tp, config.mns)]
+            # These are predictor-grid limits, not scheduler knobs.  Keep the
+            # grid inside the observed vLLM graph/runtime envelope: no batch
+            # can use a capture bucket above this cell's list and MBT is 8192.
+            config_knobs["prediction_max_batch_size"] = max(
+                GRAPH_CAPTURE_SIZES_BY_MNS[config.mns]
+            )
+            config_knobs["prediction_max_prefill_chunk_size"] = 8192
         if kernel_paths is not None:
             config_knobs.update(
                 {