Capture real FMA smoke execution evidence for restart handoff

Constraint: This checkpoint records running-smoke evidence only and must not stage data, model artifacts, or tmp outputs
Rejected: Wait for the full real FMA smoke to finish before updating handoff docs | The running-state evidence is already valuable for the next session and should not be lost
Confidence: high
Scope-risk: narrow
Directive: Keep future restart notes aligned with the live smoke status and continue using explicit file staging
Tested: Re-verified real FMA smoke is running on CPU, manifests validate, and the documented no-GPU condition explains the long training phase
Not-tested: Did not wait for Epoch 1 completion, model checkpoint emission, or downstream build-index/evaluate completion

+## 2026-06-02 真实 FMA smoke 启动并进入 CPU 长训练 checkpoint
+
+完成项：
+- 已在真实本地 FMA 数据上启动端到端 smoke：
+  - 输入目录：`acr-engine/data/raw/fma_small_audio`
+  - 输出目录：`/tmp/fma_real_smoke_stopcheck`
+- 已确认真实 FMA manifest 生成并通过校验：
+  - `catalog_references=8000`
+  - `train_queries=6401`
+  - `test_queries=1593`
+  - `val_queries=0`
+- 已确认当前环境无可用 NVIDIA GPU：
+  - `nvidia-smi` 返回 `NO_NVIDIA_GPU`
+  - `torch.cuda.is_available() = false`
+- 已确认真实 smoke 当前处于 CPU 训练阶段，且持续推进：
+  - 训练命令输出目录：`/tmp/fma_real_smoke_stopcheck/fma_models_smoke`
+  - 当前 checkpoint 时进度已推进到 `Epoch 1 step 836/3201`
+
+验证结果：
+- `check-local-ready fma ...` => `ready_for_smoke=true`
+- `validate-splits /tmp/fma_real_smoke_stopcheck/fma/manifests` => `ok=true`
+- `train.py` 进程仍在运行，`ELAPSED≈08:22`
+
+结论：
+- 真实 FMA 数据已经不只是“可检查”，而是已进入真实端到端 smoke 执行。
+- 当前慢的主因是“无 GPU + 真实 FMA 规模较大”，不是流程卡死。
+- `fma_models_smoke` 目录暂时无文件是正常现象；按当前 `train.py` 逻辑，`best_model.pt` 会在 `Epoch 1` 结束后首次落盘。
+
 ## 2026-06-02 Python 缓存噪音忽略规则补齐 checkpoint
 
 完成项：

@@ -103,3 +103,10 @@ cd /workspace/acr-engine
 
 - 已新增 `acr-engine/scripts/business_export_offline_smoke.py`，并拿到端到端离线 smoke fresh evidence。
 - 已确认链路：业务导出样例 -> 规范化 -> 项目 manifest -> `train.py --dry-run`。
+
+- 已补记真实 FMA smoke 的进行中 fresh evidence：
+  - `fma_small_audio` 已 `ready_for_smoke=true`
+  - 真实 smoke 输出目录：`/tmp/fma_real_smoke_stopcheck`
+  - manifest 校验通过：`catalog_references=8000`, `train_queries=6401`, `test_queries=1593`
+  - 当前环境无 GPU，真实 smoke 正在 CPU 上进入长训练阶段
+  - 训练中途 `fma_models_smoke/` 为空是正常现象，因为 `train.py` 在 `Epoch 1` 结束后才首次保存 `best_model.pt`

@@ -34,6 +34,23 @@
 - 跑真实开放数据 smoke
 - 继续优化准确率，尤其是 `confused` / `humming_like`
 
+### 最新真实 FMA 运行事实（2026-06-02 补记）
+
+- `fma_small.zip` 已完整落地，并已解压到 `acr-engine/data/raw/fma_small_audio`
+- `check-local-ready fma ...` 已验证：
+  - `ready_for_smoke=true`
+  - `num_audio_files=8000`
+  - `eligible_query_files=7994`
+- 真实 FMA smoke 已实际启动到：
+  - 输出目录：`/tmp/fma_real_smoke_stopcheck`
+  - manifest 校验：`ok=true`
+  - 当前训练规模：`catalog_references=8000`, `train_queries=6401`, `test_queries=1593`
+- 当前环境无 GPU：
+  - `nvidia-smi` => `NO_NVIDIA_GPU`
+  - `torch.cuda.is_available() = false`
+- 因此本轮真实 smoke 当前表现为 **CPU 长训练**，不是异常卡死。
+- 重要：`train.py` 采用 **epoch-end save** 策略，`best_model.pt` 会在 `Epoch 1` 结束后首次落盘；所以训练中途看到空的 `fma_models_smoke/` 目录是正常现象。
+
 ---
 
 ## 1. 项目是什么
@@ -734,4 +751,3 @@ seed123 最终结论：
   - `high_energy`: `3 / 1.0 / 1.0`
   - winner: `hybrid`
 - 当前第二个 bucket `prefix_000_b` 仍在继续执行
-