Add batch-aware profile and exact trace preparation

runs/frontier-qwen30-vllm020-profile-v1/freeze_frontier_profiles.py

@@ -638,11 +638,17 @@ def main() -> None:
         args.output / "moe.csv",
         args.output / "allreduce.json",
     ]
+    batch_composition_augmented = len(args.attention) > 3
     manifest = {
-        "schema_version": "frontier_qwen30_vllm020_frozen_profile.v2",
+        "schema_version": (
+            "frontier_qwen30_vllm020_frozen_profile.v3"
+            if batch_composition_augmented
+            else "frontier_qwen30_vllm020_frozen_profile.v2"
+        ),
         "profile_id": (
             "qwen3-30b-a3b-bf16-vllm020-h20-tp1-2-4-"
             "fused-mixed-total-conserving"
+            + ("-pure-prefill-batch-composition" if batch_composition_augmented else "")
         ),
         "environment_contract": {
             "hardware": "NVIDIA H20",
@@ -674,6 +680,13 @@ def main() -> None:
                 "ratio, with projected prefill + decode exactly equal to the fused total; "
                 "the split is a schema compatibility attribution, not an observation"
             ),
+            "attention_pure_prefill_batch_composition": (
+                "Direct FA3 measurements for 2/4 requests at query length 2048 and "
+                "2/4/8/16 requests at query length 512 for each TP; included only "
+                "when batch-composition attention inputs are supplied"
+                if batch_composition_augmented
+                else "not included"
+            ),
             "moe": (
                 "Replicated gate and fused top-k plus TP-local modular expert kernel; "
                 "expert measurement already includes prepare/finalize so shuffling is zero"