feat: gate prefix-ordered consumption core (algo #6)

CE-Gate 语义修订获批:块序贯 → 阶梯序前缀逐对序贯。新增 GateSpec/_UnitSlot/ _GateRun 数据结构与 _advance_prefix 纯逻辑(乱序到达下统计严格按预声明阶梯序消费,INFRA 剔除后重判防 continue 悬置,过线即冻结)。旧块路径共存,Task 6 删。
2026-07-16 23:44:32 -04:00
parent b02db74237
commit 21c360bc87
2 changed files with 269 additions and 0 deletions
@@ -793,3 +793,138 @@ async def validate_skill_local(
            shutil.rmtree(cand_dir)
        except OSError as e:
            logger.warning("候选临时目录清理失败 {}: {}", cand_dir, e)
+
+
+# ---------------------------------------------------------------------------
+# 连续并发 gate:数据结构 + 前缀消费(algo #6 语义修订:块序贯 → 阶梯序前缀逐对序贯)
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class GateSpec:
+    """单题型 gate 验证规格(runner 装配阶段产物,调度器输入)。
+
+    字段:
+        task_type: 题型。
+        target_file: 解析后生效 skill 文件名(候选物化写此文件)。
+        candidate_content: 候选 skill 全文。
+        base_skill_content: 基线侧生效 skill 全文(skill_hash 作缓存键)。
+        units: 阶梯序单元列表(已排除案例单元、截断 gate_n_max)。
+        gate_run_prefix: run_id 前缀,必须含 "_gate_"(防泄露过滤依赖)。
+    """
+
+    task_type: str
+    target_file: str
+    candidate_content: str
+    base_skill_content: str
+    units: list[QuestionUnit]
+    gate_run_prefix: str
+
+
+@dataclass
+class _UnitSlot:
+    """单个阶梯单元的双臂到达状态。
+
+    base 为单元级对错(AR pair 已折叠);cand_per_q 为逐题对错(折叠交给消费时,
+    以复用 unit_correctness_view 并保留逐题溯源)。INFRA 标志与结果互斥。
+    """
+
+    unit: QuestionUnit
+    base: bool | None = None
+    cand_per_q: dict[str, bool] | None = None
+    base_infra: bool = False
+    cand_infra: bool = False
+
+    def resolved(self) -> bool:
+        """双臂均已出结果(含 INFRA 判定)。"""
+        base_done = self.base is not None or self.base_infra
+        cand_done = self.cand_per_q is not None or self.cand_infra
+        return base_done and cand_done
+
+    def excluded(self) -> bool:
+        """任一臂 INFRA 即整单元剔除(不入配对)。"""
+        return self.base_infra or self.cand_infra
+
+
+@dataclass
+class _GateRun:
+    """单题型 gate 的运行时状态(计数器 + 前缀指针 + 证据)。"""
+
+    spec: GateSpec
+    slots: list[_UnitSlot]
+    s_hash: str
+    prefix_ptr: int = 0
+    w: int = 0
+    l: int = 0  # noqa: E741
+    n_used: int = 0
+    n_excluded: int = 0
+    errors: int = 0
+    infra_denom: int = 0
+    frozen: bool = False
+    verdict: GateVerdict | None = None
+    base_obs: dict[str, bool] = field(default_factory=dict)
+    cand_obs: dict[str, bool] = field(default_factory=dict)
+    candidate_per_q: dict[str, bool] = field(default_factory=dict)
+    evidence_rows: list[dict] = field(default_factory=list)
+
+    @classmethod
+    def from_spec(cls, spec: GateSpec) -> _GateRun:
+        """由规格构造初始状态(slots 与阶梯序一一对应)。"""
+        return cls(
+            spec=spec,
+            slots=[_UnitSlot(unit=u) for u in spec.units],
+            s_hash=skill_hash(spec.base_skill_content),
+        )
+
+
+def _advance_prefix(run: _GateRun, params: GateParams) -> None:
+    """沿阶梯序消费"已配齐前缀",逐单元更新 (W,L) 并判定,过线即冻结。
+
+    统计合法性关键(设计 v3 §1 / Codex C1):严禁按到达序消费——base 臂缓存命中
+    瞬间返回、cand 臂必新鲜跑,两臂延迟不对称,若 cand 延迟与对错相关,早到翻转
+    对系统性偏向 W 型 → e-值虚高假接受。前缀消费把判定顺序钉回预声明阶梯序,
+    anytime-valid 无条件成立;INFRA 单元视为"已解决(剔除)"不阻塞前缀。
+    """
+    while not run.frozen and run.prefix_ptr < len(run.slots):
+        slot = run.slots[run.prefix_ptr]
+        if not slot.resolved():
+            return
+        rank = run.prefix_ptr
+        run.prefix_ptr += 1
+        if slot.excluded():
+            run.n_excluded += 1
+            # 剔除使 n_remaining 缩小,必须重判(Codex plan 审 C1):否则尾部全 INFRA
+            # 时 verdict 停留在 "continue",绕过题尽第四出口且 _finalize_outcome
+            # 查 stop_reason 映射 KeyError。n_used==0(纯前导 INFRA)时无证据可判,跳过。
+            if run.n_used > 0:
+                n_remaining = (len(run.slots) - run.n_excluded) - run.n_used
+                run.verdict = gate_decision(run.w, run.l, run.n_used, n_remaining, params=params)
+                if run.verdict.decision != "continue":
+                    run.frozen = True
+            continue
+        uid = slot.unit.unit_id
+        assert slot.base is not None and slot.cand_per_q is not None
+        c_units = unit_correctness_view([slot.unit], slot.cand_per_q)
+        pair_result = pair_block({uid: slot.base}, c_units, [uid])
+        run.candidate_per_q.update(slot.cand_per_q)
+        for u, (b, c) in pair_result.observed.items():
+            run.base_obs[u] = b
+            run.cand_obs[u] = c
+        run.w += pair_result.w
+        run.l += pair_result.l
+        run.n_used += 1
+        n_remaining = (len(run.slots) - run.n_excluded) - run.n_used
+        run.verdict = gate_decision(run.w, run.l, run.n_used, n_remaining, params=params)
+        run.evidence_rows.append(
+            {
+                "question_id": uid,
+                "task_type": run.spec.task_type,
+                "ladder_rank": rank,
+                "baseline_correct": slot.base,
+                "candidate_correct": c_units[uid],
+                "e_value": run.verdict.e_value,
+                "stop_reason": "",
+            }
+        )
+        if run.verdict.decision != "continue":
+            run.frozen = True