层1/T2: teacher.py 批量生成 + sha256 JSONL 缓存；teacher 改定 MiniMax-M3

- teacher.py: 通用 OpenAI 兼容客户端（配置驱动 base_url，替代 OpenRouter 专用）；
  缓存即断点（逐条落盘+flush，重跑自动续传）；单条失败先落盘其余、结束汇总显式报错；
  M3 思考段 <think>...</think> 入库前剥离（只剥开头一段）
- configs.py: 新增 TeacherGenConfig（采样参数显式化；连接三元组走 .env）
- scripts/generate_teacher_completions.py: 自包含生成脚本（本地跑，与训练侧
  同 seed 同子集约束已注明）
- teacher 决策变更同步：.env.example / docs/00 关键设定与存档点 / docs/02
- tests/test_teacher.py: 10 个单测（假客户端注入），含与 attach 的端到端契约闭环

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/test_teacher.py

@@ -0,0 +1,143 @@
+"""层 1 / T2：teacher 批量生成与缓存单测。
+
+用假 OpenAI 客户端注入（TeacherClient 的测试口），验证思考段剥离、缓存契约
+（与 data.attach_teacher_completions 的端到端闭环）、断点续传、失败汇总。
+"""
+
+from types import SimpleNamespace
+
+import pytest
+from datasets import Dataset
+
+from ars_opd.configs import TeacherGenConfig
+from ars_opd.data import attach_teacher_completions, prompt_key
+from ars_opd.teacher import TeacherClient, _load_teacher_env, generate_completions
+
+
+class FakeClient:
+    """最小 OpenAI 客户端替身：chat.completions.create 按 responder 出内容。"""
+
+    def __init__(self, responder):
+        self.calls = []
+        self._responder = responder
+        self.chat = SimpleNamespace(completions=SimpleNamespace(create=self._create))
+
+    def _create(self, model, messages, **kwargs):
+        self.calls.append(messages)
+        content = self._responder(messages)
+        return SimpleNamespace(
+            choices=[SimpleNamespace(message=SimpleNamespace(content=content))]
+        )
+
+
+def make_teacher(responder, **cfg_overrides):
+    cfg = TeacherGenConfig(**cfg_overrides)
+    return TeacherClient(cfg, client=FakeClient(responder), model="fake-m3")
+
+
+def user(q):
+    return [{"role": "user", "content": q}]
+
+
+# ---------------------------------------------------------------------------
+# TeacherClient.generate
+# ---------------------------------------------------------------------------
+
+
+def test_剥离开头思考段():
+    teacher = make_teacher(lambda m: "<think>心算一下</think>\n答案是 42")
+    assert teacher.generate(user("q")) == "答案是 42"
+
+
+def test_正文中的think字样不误删():
+    teacher = make_teacher(lambda m: "<think>x</think>正文提到 <think> 标签本身")
+    assert teacher.generate(user("q")) == "正文提到 <think> 标签本身"
+
+
+def test_只剩思考段等于空解答_报错():
+    teacher = make_teacher(lambda m: "<think>思考被截断在半途")
+    # 未闭合的 think 段剥不掉，但闭合后为空的要报错
+    teacher_empty = make_teacher(lambda m: "<think>只有思考</think>  ")
+    with pytest.raises(ValueError, match="空解答"):
+        teacher_empty.generate(user("q"))
+    # 未闭合时保留原文（宁可保留可疑内容也不静默删成空）
+    assert "<think>" in teacher.generate(user("q"))
+
+
+def test_关闭strip_think则原样保留():
+    teacher = make_teacher(lambda m: "<think>a</think>b", strip_think=False)
+    assert teacher.generate(user("q")) == "<think>a</think>b"
+
+
+def test_system_prompt前置():
+    teacher = make_teacher(lambda m: "ok", system_prompt="你是数学助教")
+    teacher.generate(user("q"))
+    sent = teacher.client.calls[0]
+    assert sent[0] == {"role": "system", "content": "你是数学助教"}
+    assert sent[1]["role"] == "user"
+
+
+def test_注入client但不给model报错():
+    with pytest.raises(ValueError, match="model"):
+        TeacherClient(TeacherGenConfig(), client=FakeClient(lambda m: "x"), model=None)
+
+
+# ---------------------------------------------------------------------------
+# generate_completions：缓存契约与断点续传
+# ---------------------------------------------------------------------------
+
+
+def test_端到端契约_生成的缓存能被attach消费(tmp_path):
+    cache = str(tmp_path / "cache.jsonl")
+    prompts = [user("1+1=?"), user("2+2=?")]
+    teacher = make_teacher(lambda m: f"对「{m[-1]['content']}」的解答")
+
+    generate_completions(prompts, cache, teacher)
+
+    ds = Dataset.from_list([{"messages": p} for p in prompts])
+    out = attach_teacher_completions(ds, cache)
+    assert out[0]["messages"][-1]["content"] == "对「1+1=?」的解答"
+    assert out[1]["messages"][-1]["content"] == "对「2+2=?」的解答"
+
+
+def test_断点续传_已缓存的不重新生成(tmp_path):
+    cache = str(tmp_path / "cache.jsonl")
+    prompts = [user("q1"), user("q2")]
+    teacher = make_teacher(lambda m: "a")
+
+    generate_completions([prompts[0]], cache, teacher)
+    assert len(teacher.client.calls) == 1
+    generate_completions(prompts, cache, teacher)  # q1 命中缓存
+    assert len(teacher.client.calls) == 2  # 只多了 q2 一次调用
+
+
+def test_单条失败_其余落盘_结束时汇总报错(tmp_path):
+    cache = str(tmp_path / "cache.jsonl")
+    prompts = [user("好题"), user("坏题")]
+
+    def responder(m):
+        if m[-1]["content"] == "坏题":
+            raise RuntimeError("网关 500")
+        return "解答"
+
+    teacher = make_teacher(responder)
+    with pytest.raises(RuntimeError, match="1/2"):
+        generate_completions(prompts, cache, teacher)
+
+    # 成功的那条已经在缓存里，重跑只会补坏题
+    from ars_opd.teacher import _cached_keys
+    from pathlib import Path
+
+    assert _cached_keys(Path(cache)) == {prompt_key(prompts[0])}
+
+
+# ---------------------------------------------------------------------------
+# .env 读取
+# ---------------------------------------------------------------------------
+
+
+def test_env缺失显式报错(monkeypatch):
+    for name in ("TEACHER_API_BASE", "TEACHER_API_KEY", "TEACHER_MODEL"):
+        monkeypatch.delenv(name, raising=False)
+    with pytest.raises(ValueError, match="TEACHER_API_BASE"):
+        _load_teacher_env(env_file="/不存在的路径/.env")