fix(evals): preserve nested lm-eval artifacts for sweep summaries

benchmarks/benchmark_lib.sh

@@ -2163,7 +2163,12 @@ append_lm_eval_summary() {
     fi
 
     # Copy the complete allowlisted eval artifact set before removing its temp dir.
-    stage_eval_artifacts "$(pwd)" "$out_dir" || return $?
+    local artifact_dir
+    local artifact_sources=("$out_dir")
+    while IFS= read -r -d '' artifact_dir; do
+        [ "$artifact_dir" = "$out_dir" ] || artifact_sources+=("$artifact_dir")
+    done < <(find "$out_dir" -type d -print0 2>/dev/null)
+    stage_eval_artifacts "$(pwd)" "${artifact_sources[@]}" || return $?
 
     # Best-effort cleanup of the temp directory
     if [ -n "${out_dir}" ] && [ -d "${out_dir}" ]; then

docs/eval-agentx-procedures.md

@@ -171,6 +171,8 @@ gh run download "$RUN_ID" --repo SemiAnalysisAI/InferenceX \
 
 Retain `meta_env.json`, `results*.json`, and `sample*.jsonl`. Agentic SWE-bench additionally uploads `agent_preds.json`, `predictions.jsonl`, `swebench_report_*.json`, and trajectory files in the single-node template. The aggregate is a navigation aid, not a substitute for raw samples and batch completeness.
 
+Single-concurrency lm-eval can write `results*.json` and `sample*.jsonl` below a model-named subdirectory of `EVAL_RESULT_DIR`. `append_lm_eval_summary` recursively stages only the allowlisted eval artifacts into the workspace root before removing the temporary directory, where the workflow's flat upload patterns can find them.
+
 ## 7. Run AgentX: fast feedback versus canonical evidence
 
 AgentX is AIPerf `inferencex-agentx-mvp` trace replay, not a fixed-token synthetic benchmark. The checked-in default uses ten additional warmup requests per trajectory lane and the recipe's configured profile duration. `agentx-fast` forces one warmup request per lane and a 1,200-second profile. It affects single- and multi-node AgentX throughput only. Fixed-sequence throughput and evals remain canonical. Fast runs are not eligible for artifact reuse ([workflow policy](../.github/workflows/README.md#agentx-fast-mode), [fast replay settings](../benchmarks/benchmark_lib.sh#L2104-L2128)).

docs/eval-agentx-procedures_zh.md

@@ -171,6 +171,8 @@ gh run download "$RUN_ID" --repo SemiAnalysisAI/InferenceX \
 
 保留 `meta_env.json`、`results*.json` 和 `sample*.jsonl`。Agentic SWE-bench 在单节点模板中还会上传 `agent_preds.json`、`predictions.jsonl`、`swebench_report_*.json` 和 trajectory 文件。Aggregate 是导航工具，不能替代原始样本与 batch 完整性证据。
 
+单并发 lm-eval 可能会把 `results*.json` 和 `sample*.jsonl` 写入 `EVAL_RESULT_DIR` 下以模型命名的子目录。`append_lm_eval_summary` 会在删除临时目录前，递归地把允许列表中的 eval artifact 暂存到 workspace 根目录，使工作流的扁平上传模式能够找到这些文件。
+
 ## 7. 运行 AgentX：快速反馈与 canonical 证据
 
 AgentX 是 AIPerf `inferencex-agentx-mvp` trace replay，不是固定 token 的合成 benchmark。仓库默认设置对每条 trajectory lane 额外执行十个 warmup 请求，并使用 recipe 配置的 profile 时长。`agentx-fast` 强制每条 lane 只运行一个 warmup 请求，并将 profile 设为 1,200 秒。它只影响单节点和多节点 AgentX 吞吐量；定长序列吞吐量与 eval 保持 canonical。Fast 运行不符合 artifact reuse 条件（[工作流策略](../.github/workflows/README.md#agentx-fast-mode)、[fast replay 设置](../benchmarks/benchmark_lib.sh#L2104-L2128)）。

utils/evals/test_run_eval_dispatch.py

@@ -1551,6 +1551,42 @@ def test_summary_stages_bfcl_upstream_archive_before_cleanup(tmp_path: Path) ->
     assert not results_dir.exists()
 
 
+def test_summary_stages_nested_lm_eval_outputs_before_cleanup(tmp_path: Path) -> None:
+    work_dir = tmp_path / "work"
+    results_dir = tmp_path / "results"
+    nested_results_dir = results_dir / "local-chat-completions__test-model"
+    work_dir.mkdir()
+    nested_results_dir.mkdir(parents=True)
+    result = nested_results_dir / "results_2026-09-09T12-00-00.json"
+    sample = nested_results_dir / "samples_gsm8k_2026-09-09T12-00-00.jsonl"
+    result_content = '{"lm_eval_version": "test"}'
+    sample_content = '{"doc_id": 1}\n'
+    result.write_text(result_content)
+    sample.write_text(sample_content)
+    (nested_results_dir / "debug.log").write_text("do not stage")
+    script = r"""
+source "$BENCHMARK_LIB"
+cd "$WORK_DIR"
+append_lm_eval_summary >/dev/null
+"""
+    env = {
+        **os.environ,
+        "BENCHMARK_LIB": str(BENCHMARK_LIB),
+        "WORK_DIR": str(work_dir),
+        "EVAL_RESULT_DIR": str(results_dir),
+        "MODEL": "test-model",
+        "CONC": "7",
+        "KV_OFFLOADING": "none",
+    }
+
+    subprocess.run(["bash", "-c", script], env=env, check=True)
+
+    assert (work_dir / result.name).read_text() == result_content
+    assert (work_dir / sample.name).read_text() == sample_content
+    assert not (work_dir / "debug.log").exists()
+    assert not results_dir.exists()
+
+
 def test_stage_eval_artifacts_copies_eval_outputs_only(tmp_path: Path) -> None:
     source_one = tmp_path / "source-one"
     source_two = tmp_path / "source-two"