fix: early_stop patience counts epochs not steps

app/harness/checkpoint.py

@@ -94,7 +94,7 @@ def serialize_state(state: Any) -> dict[str, Any]:
         "eval_prev_run_id": state.eval_prev_run_id,
         "baseline_skills_version": state.baseline_skills_version,
         "baseline_prompts_version": state.baseline_prompts_version,
-        "steps_since_best_improved": state.steps_since_best_improved,
+        "epochs_since_best_improved": state.epochs_since_best_improved,
         "epoch_start_skills": state.epoch_start_skills,
         "changed_task_types_this_epoch": sorted(state.changed_task_types_this_epoch),
         "rejected_buffer": {k: [asdict(x) for x in v] for k, v in state.rejected_buffer.items()},
@@ -145,7 +145,7 @@ def deserialize_state_fields(d: dict[str, Any]) -> dict[str, Any]:
         "eval_prev_run_id": d["eval_prev_run_id"],
         "baseline_skills_version": d["baseline_skills_version"],
         "baseline_prompts_version": d["baseline_prompts_version"],
-        "steps_since_best_improved": d["steps_since_best_improved"],
+        "epochs_since_best_improved": d["epochs_since_best_improved"],
         "epoch_start_skills": d["epoch_start_skills"],
         "changed_task_types_this_epoch": set(d["changed_task_types_this_epoch"]),
         "rejected_buffer": {

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}",

tests/integration/test_checkpoint_pair.py

@@ -90,7 +90,7 @@ class _FakeState:
     eval_prev_run_id: str = "run-0"
     baseline_skills_version: str = "v1"
     baseline_prompts_version: str = "v1"
-    steps_since_best_improved: int = 0
+    epochs_since_best_improved: int = 0
     epoch_start_skills: str = "v1"
     changed_task_types_this_epoch: set[str] = field(default_factory=set)
     rejected_buffer: dict = field(default_factory=dict)

tests/unit/test_harness_checkpoint.py

@@ -116,7 +116,7 @@ class _FakeState:
     eval_prev_run_id: str
     baseline_skills_version: str
     baseline_prompts_version: str
-    steps_since_best_improved: int
+    epochs_since_best_improved: int
     epoch_start_skills: str
     changed_task_types_this_epoch: set[str]
     rejected_buffer: dict[str, list[RejectedEdit]]
@@ -135,7 +135,7 @@ def _make_state() -> _FakeState:
         eval_prev_run_id="run-abc",
         baseline_skills_version="v1",
         baseline_prompts_version="v1",
-        steps_since_best_improved=2,
+        epochs_since_best_improved=2,
         epoch_start_skills="v1",
         changed_task_types_this_epoch={"temporal", "causal"},
         rejected_buffer={"temporal": [_make_rejected_edit()]},
@@ -201,7 +201,7 @@ class TestSerializeDeserializeRoundtrip:
         assert restored["eval_prev_run_id"] == state.eval_prev_run_id
         assert restored["baseline_skills_version"] == state.baseline_skills_version
         assert restored["baseline_prompts_version"] == state.baseline_prompts_version
-        assert restored["steps_since_best_improved"] == state.steps_since_best_improved
+        assert restored["epochs_since_best_improved"] == state.epochs_since_best_improved
         assert restored["epoch_start_skills"] == state.epoch_start_skills
         assert restored["changed_task_types_this_epoch"] == state.changed_task_types_this_epoch
         assert restored["gate_cooldown"] == state.gate_cooldown

tests/unit/test_harness_runner.py

@@ -349,92 +349,56 @@ class TestBuildComparisonPairs:
 # =========================================================================
 
 
+def _write_manifest_with_best(tmp_path: Path, best_epoch: int) -> None:
+    """写含 best.epoch 的 manifest，供 _should_early_stop 读 read_best。"""
+    manifest = {
+        "name": "test",
+        "store": ".",
+        "current": {"videos": "v", "questions": "q", "skills": "s/v1", "prompts": "p/v1"},
+        "best": {"epoch": best_epoch, "val_acc": 0.5},
+        "history": [],
+    }
+    (tmp_path / "manifest.json").write_text(json.dumps(manifest))
+
+
 class TestShouldEarlyStop:
-    """_should_early_stop 步粒度 early stop。"""
+    """_should_early_stop epoch 粒度 early stop（patience 以 epoch 计）。"""
 
     def test_improved_this_epoch_resets(self, tmp_path: Path) -> None:
         """本 epoch best 刷新时重置计数器。"""
-        # 写 manifest + best
-        manifest = {
-            "name": "test",
-            "store": ".",
-            "current": {"videos": "v", "questions": "q", "skills": "s/v1", "prompts": "p/v1"},
-            "best": {"epoch": 2, "val_acc": 0.9},
-            "history": [],
-        }
-        (tmp_path / "manifest.json").write_text(json.dumps(manifest))
+        _write_manifest_with_best(tmp_path, best_epoch=2)
 
         state = MagicMock()
-        state.steps_since_best_improved = 10
+        state.epochs_since_best_improved = 3
 
-        result = _should_early_stop(tmp_path, epoch=2, steps_this_epoch=5, state=state, patience=20)
+        result = _should_early_stop(tmp_path, epoch=2, state=state, patience=2)
         assert result is False
-        assert state.steps_since_best_improved == 0
-
-    def test_no_improvement_accumulates(self, tmp_path: Path) -> None:
-        """未刷新时累加步数。"""
-        manifest = {
-            "name": "test",
-            "store": ".",
-            "current": {"videos": "v", "questions": "q", "skills": "s/v1", "prompts": "p/v1"},
-            "best": {"epoch": 1, "val_acc": 0.5},
-            "history": [],
-        }
-        (tmp_path / "manifest.json").write_text(json.dumps(manifest))
-
-        state = MagicMock()
-        state.steps_since_best_improved = 15
-
-        result = _should_early_stop(tmp_path, epoch=3, steps_this_epoch=5, state=state, patience=20)
-        assert result is True  # 15 + 5 = 20 >= 20
-        assert state.steps_since_best_improved == 20
+        assert state.epochs_since_best_improved == 0
 
     def test_below_patience_continues(self, tmp_path: Path) -> None:
-        """累计步数未达阈值时继续。"""
-        manifest = {
-            "name": "test",
-            "store": ".",
-            "current": {"videos": "v", "questions": "q", "skills": "s/v1", "prompts": "p/v1"},
-            "best": {"epoch": 1, "val_acc": 0.5},
-            "history": [],
-        }
-        (tmp_path / "manifest.json").write_text(json.dumps(manifest))
+        """未达 patience 个 epoch 无刷新时继续。"""
+        _write_manifest_with_best(tmp_path, best_epoch=1)
 
         state = MagicMock()
-        state.steps_since_best_improved = 10
+        state.epochs_since_best_improved = 0
 
-        result = _should_early_stop(tmp_path, epoch=3, steps_this_epoch=5, state=state, patience=20)
+        result = _should_early_stop(tmp_path, epoch=2, state=state, patience=3)
         assert result is False
-        assert state.steps_since_best_improved == 15
+        assert state.epochs_since_best_improved == 1
 
-    def test_step_granularity(self, tmp_path: Path) -> None:
-        """步粒度而非 epoch 粒度。"""
-        manifest = {
-            "name": "test",
-            "store": ".",
-            "current": {"videos": "v", "questions": "q", "skills": "s/v1", "prompts": "p/v1"},
-            "best": {"epoch": 1, "val_acc": 0.5},
-            "history": [],
-        }
-        (tmp_path / "manifest.json").write_text(json.dumps(manifest))
+    def test_early_stop_counts_epochs_not_steps(self, tmp_path: Path) -> None:
+        """patience=2 表示连续 2 个 epoch 无 best 刷新才停（不是步数）。"""
+        _write_manifest_with_best(tmp_path, best_epoch=1)
 
         state = MagicMock()
-        # 连续 3 个 epoch，每个 3 步
-        state.steps_since_best_improved = 0
-        for ep in range(2, 5):
-            stopped = _should_early_stop(
-                tmp_path, epoch=ep, steps_this_epoch=3, state=state, patience=10
-            )
-            if ep < 4:
-                assert stopped is False
-            else:
-                # 3+3+3=9 < 10 但第三轮后 9+3=12>=10 在 ep=5 触发
-                # 实际：ep=2 → 3, ep=3 → 6, ep=4 → 9
-                assert stopped is False
-        stopped = _should_early_stop(
-            tmp_path, epoch=5, steps_this_epoch=3, state=state, patience=10
-        )
-        assert stopped is True  # 9+3=12>=10
+        state.epochs_since_best_improved = 0
+
+        # epoch 2 无刷新 → 1 → 不停
+        assert _should_early_stop(tmp_path, epoch=2, state=state, patience=2) is False
+        assert state.epochs_since_best_improved == 1
+        # epoch 3 无刷新 → 2 → 停
+        assert _should_early_stop(tmp_path, epoch=3, state=state, patience=2) is True
+        assert state.epochs_since_best_improved == 2
 
 
 # =========================================================================
@@ -667,7 +631,7 @@ class TestTrainState:
         assert state.system_packs == []
         assert state.tool_packs == []
         assert state.changed_task_types_this_epoch == set()
-        assert state.steps_since_best_improved == 0
+        assert state.epochs_since_best_improved == 0
 
 
 # =========================================================================