Prepare sealed exact-trace real smoke

runs/frontier-fidelity-envelope-v1/run_qwen30_exact_trace_real_anchor.sh

@@ -39,6 +39,7 @@ REQUEST_COUNT="$(wc -l < "${REQUESTS_FILE}")"
 echo "QWEN30_EXACT_TRACE_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 trace=${TRACE_LABEL} requests=${REQUEST_COUNT} source=${REQUESTS_FILE} arrivals=original600s input_output_prompt=exact prefix=on block=16 tpot_slo=150ms output=${OUTPUT_ROOT} expected_wall=12-35m hard_wall=3600s"
 date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
 sha256sum qwen30_exact_trace_client.py run_qwen30_exact_trace_real_anchor.sh \
+  ../frontier-phase-factorial-v0/qwen30_prefill_client.py \
   > "${OUTPUT_ROOT}/provenance/source.sha256"
 sha256sum "${REQUESTS_FILE}" > "${OUTPUT_ROOT}/provenance/requests.sha256"
 sha256sum "${MODEL_ROOT}/config.json" > "${OUTPUT_ROOT}/provenance/model.sha256"
@@ -77,6 +78,14 @@ if [[ "${READY}" -ne 1 ]]; then
   exit 1
 fi
 
+# Warm execution kernels with four prefix-disjoint requests, then start the
+# measured replay from an empty scheduler queue.  The unrelated 2K-token KV
+# footprint remains explicit and negligible relative to the configured cache.
+"${VENV_ROOT}/bin/python" ../frontier-phase-factorial-v0/qwen30_prefill_client.py \
+  --port "${SERVER_PORT}" --served-model "${SERVED_MODEL}" \
+  --model-path "${MODEL_ROOT}" --rate 1 --requests 4 --input-tokens 512 \
+  --output-tokens 1 --output "${OUTPUT_ROOT}/results/warmup.json"
+
 "${VENV_ROOT}/bin/python" qwen30_exact_trace_client.py \
   --port "${SERVER_PORT}" --requests-file "${REQUESTS_FILE}" \
   --output "${OUTPUT_ROOT}/results/result.json" --tpot-slo-ms 150 \