fix: early_stop patience counts epochs not steps

app/harness/runner.py

@@ -115,7 +115,7 @@ class _TrainState:
     global_step: int = 0
     changed_task_types_this_epoch: set[str] = field(default_factory=set)
     epoch_start_skills: dict[str, str] = field(default_factory=dict)
-    steps_since_best_improved: int = 0
+    epochs_since_best_improved: int = 0
     gate_epoch_observed: bool = False
     probations: dict[str, Probation] = field(default_factory=dict)
     gate_cooldown: dict[str, int] = field(default_factory=dict)
@@ -291,18 +291,16 @@ def _snapshot_current_skills(skills_dir: Path) -> dict[str, str]:
 def _should_early_stop(
     workspace_dir: Path,
     epoch: int,
-    steps_this_epoch: int,
     state: _TrainState,
     patience: int,
 ) -> bool:
-    """步粒度 early stop：本 epoch best 未刷新则累加本 epoch 步数。
+    """epoch 粒度 early stop：本 epoch best 未刷新则计数 +1。
 
     参数:
         workspace_dir: workspace 目录（读 manifest best）。
         epoch: 当前 epoch。
-        steps_this_epoch: 本 epoch 的 step 总数。
-        state: 训练状态（steps_since_best_improved 就地更新）。
-        patience: early_stop_patience。
+        state: 训练状态（epochs_since_best_improved 就地更新）。
+        patience: early_stop_patience（连续无刷新的 epoch 数上限）。
 
     返回:
         是否触发 early stop。
@@ -310,10 +308,10 @@ def _should_early_stop(
     best = read_best(workspace_dir)
     improved_this_epoch = best is not None and best.get("epoch") == epoch
     if improved_this_epoch:
-        state.steps_since_best_improved = 0
+        state.epochs_since_best_improved = 0
         return False
-    state.steps_since_best_improved += steps_this_epoch
-    return state.steps_since_best_improved >= patience
+    state.epochs_since_best_improved += 1
+    return state.epochs_since_best_improved >= patience
 
 
 def _compute_total_steps(pools: Pools, correctness: dict[str, bool], config: RunConfig) -> int:
@@ -849,7 +847,6 @@ class Runner:
             if _should_early_stop(
                 self._config.workspace_dir,
                 epoch,
-                len(batches),
                 state,
                 self._config.early_stop_patience,
             ):
@@ -1588,7 +1585,7 @@ class Runner:
         state.best_val_acc = eval_acc
         state.best_skills_version = skills_v
         state.best_prompts_version = prompts_v
-        state.steps_since_best_improved = 0
+        state.epochs_since_best_improved = 0
         update_best(
             self._config.workspace_dir,
             skills=f"skills/{skills_v}",