feat(pipeline): replace QuestionFamilySpec with TaskTypeStrategy

- SlotAssignment: remove family field, strategy looked up at process time
- PipelineConfig: remove family_ratios field
- _assign_slots: remove family_ratios and rng params (pure deterministic)
- _process_one_slot: use get_strategy() for sampling, generation, gates
- Add sub_pattern support (level/constraint override, instruction injection)
- Add strategy.extra_gates() check after standard gates
- load_pipeline_config: stop reading family_ratios from YAML
- Update tools/generate_questions.py seed override and dry-run log
- Update all integration tests to match new API

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

app/question_gen/pipeline_v2.py

@@ -34,11 +34,11 @@ import numpy as np
 import yaml
 from loguru import logger
 
-from app.question_gen.families import QuestionFamilySpec, get_family_for_slot
 from app.question_gen.gates import GateReport, GateResult, GateVerdict, run_gates
 from app.question_gen.generator_v2 import CandidateQuestion, generate_one_v2
 from app.question_gen.postprocess import run_postprocess
-from app.question_gen.sampler_v2 import _TASK_TYPE_TO_LEVEL, sample_material_v2
+from app.question_gen.sampler_v2 import sample_material_v2
+from app.question_gen.strategy import get_strategy
 from core.types import GeneratedQuestion
 
 if TYPE_CHECKING:
@@ -62,14 +62,12 @@ class SlotAssignment:
         slot_id: slot 唯一标识。
         video_id: 分配到的视频 ID。
         task_type: 任务类型。
-        family: 分配的问题家族规格。
-        seq: slot 序号（同 task_type 内从 1 开始）。
+        seq: slot 序号。
     """
 
     slot_id: str
     video_id: str
     task_type: str
-    family: QuestionFamilySpec
     seq: int
 
 
@@ -93,7 +91,6 @@ class PipelineConfig:
     """管线配置。
 
     属性:
-        family_ratios: 家族名到权重的映射。
         per_type: 每种 task_type 生成的题目数。
         retry_limit: 单 slot 最大重出次数。
         heavy_sample_rate: 重量抽检采样比例 [0.0, 1.0]。
@@ -103,7 +100,6 @@ class PipelineConfig:
         output_dir: 输出目录。
     """
 
-    family_ratios: dict[str, float]
     per_type: int
     retry_limit: int
     heavy_sample_rate: float
@@ -146,12 +142,7 @@ def load_pipeline_config(yaml_path: Path) -> PipelineConfig:
 
     section = raw["question_gen_v2"]
 
-    # 家族名统一为大写
-    raw_ratios = section["family_ratios"]
-    family_ratios = {k.upper(): float(v) for k, v in raw_ratios.items()}
-
     return PipelineConfig(
-        family_ratios=family_ratios,
         per_type=int(section["per_type"]),
         retry_limit=int(section["retry_limit"]),
         heavy_sample_rate=float(section["heavy_sample_rate"]),
@@ -171,20 +162,16 @@ def _assign_slots(
     video_ids: list[str],
     task_types: list[str],
     per_type: int,
-    family_ratios: dict[str, float],
-    rng: random.Random,
 ) -> list[SlotAssignment]:
     """将出题目标分配为具体 slot 列表。
 
     总 slot 数 = len(task_types) * per_type。
-    在视频间 round-robin 分配，每个 slot 通过 get_family_for_slot 决定家族。
+    在视频间 round-robin 分配。不再选择 family — strategy 在处理时查找。
 
     参数:
         video_ids: 视频 ID 列表。
         task_types: 任务类型列表。
         per_type: 每种 task_type 的目标题数。
-        family_ratios: 家族权重映射。
-        rng: 可控随机数生成器。
 
     返回:
         SlotAssignment 列表。
@@ -195,7 +182,6 @@ def _assign_slots(
     for task_type in task_types:
         for i in range(per_type):
             video_id = video_ids[i % len(video_ids)]
-            family = get_family_for_slot(task_type, family_ratios, rng)
             global_seq += 1
             slot_id = f"{task_type}_{global_seq:04d}"
             slots.append(
@@ -203,7 +189,6 @@ def _assign_slots(
                     slot_id=slot_id,
                     video_id=video_id,
                     task_type=task_type,
-                    family=family,
                     seq=global_seq,
                 )
             )
@@ -348,6 +333,9 @@ async def _process_one_slot(
     _RESAMPLE_VIDEO_INTERVAL = 1
 
     async with sem:
+        strategy = get_strategy(slot.task_type)
+        sub_pattern = strategy.select_sub_pattern(rng)
+
         prev_reason: str | None = None
         current_tree = tree
         current_video_id = slot.video_id
@@ -366,15 +354,25 @@ async def _process_one_slot(
                         current_video_id,
                     )
 
-            # Phase 1: 采样素材
+            # Phase 1: 采样素材（sub_pattern 可覆盖 level 和 constraint）
+            level = (
+                sub_pattern.sampling_level_override
+                if sub_pattern and sub_pattern.sampling_level_override is not None
+                else strategy.sampling_level
+            )
+            constraint = (
+                sub_pattern.constraint_override
+                if sub_pattern and sub_pattern.constraint_override is not None
+                else strategy.sampling_constraint
+            )
             try:
                 material = sample_material_v2(
                     tree=current_tree,
                     task_type=slot.task_type,
                     used_node_ids=used_node_ids,
                     rng=rng,
-                    level=_TASK_TYPE_TO_LEVEL[slot.task_type],
-                    constraint=slot.family.sampling,
+                    level=level,
+                    constraint=constraint,
                 )
             except (RuntimeError, KeyError) as e:
                 logger.warning(
@@ -392,11 +390,14 @@ async def _process_one_slot(
                     vlm=vlm,
                     tree=current_tree,
                     material=material,
-                    family_spec=slot.family,
                     task_type=slot.task_type,
                     seq=slot.seq,
                     video_id=current_video_id,
+                    prompt_template=strategy.prompt_template,
+                    strategy_name=strategy.strategy_name,
+                    skill_target=strategy.skill_target,
                     reject_reason=prev_reason,
+                    sub_pattern_instruction=sub_pattern.instruction if sub_pattern else None,
                     session_id=session_id,
                 )
             except (ValueError, FileNotFoundError, OSError, Exception) as e:
@@ -416,11 +417,12 @@ async def _process_one_slot(
                 run_id=run_id,
                 slot_id=slot.slot_id,
                 video_id=current_video_id,
-                family=slot.family.name,
+                family=strategy.strategy_name,
                 task_type=slot.task_type,
-                skill_target=slot.family.skill_target,
+                skill_target=strategy.skill_target,
                 attempt=attempt,
                 question_text=candidate.question,
+                sub_pattern=sub_pattern.name if sub_pattern else None,
             )
 
             # Phase 4: 后处理
@@ -468,7 +470,7 @@ async def _process_one_slot(
                     candidate=candidate,
                     tree=tree,
                     llm=llm,
-                    family_spec=slot.family,
+                    leak_probe_template=strategy.leak_probe_template,
                     postprocess=pp,
                     vlm=vlm,
                     session_id=session_id,
@@ -485,6 +487,21 @@ async def _process_one_slot(
                 continue
             store.update_gates(item_id, report)
 
+            # 题型专属额外 gate
+            extra_results = strategy.extra_gates(candidate)
+            if any(r.verdict == GateVerdict.FAIL for r in extra_results):
+                prev_reason = "; ".join(
+                    r.reason for r in extra_results if r.verdict == GateVerdict.FAIL
+                )
+                logger.info(
+                    "slot {} 额外 gate 失败 (attempt {}/{}): {}",
+                    slot.slot_id,
+                    attempt,
+                    config.retry_limit,
+                    prev_reason,
+                )
+                continue
+
             if not report.passed:
                 prev_reason = report.reject_reason
                 logger.info(
@@ -511,7 +528,7 @@ async def _process_one_slot(
             # Phase 8: 通过全部检查 → 接受（使用洗牌后的选项和答案）
             result = _to_generated_question(
                 candidate,
-                family=slot.family.name,
+                family=strategy.strategy_name,
                 options=pp.options,
                 answer=pp.answer,
             )
@@ -762,7 +779,7 @@ async def run_pipeline_v2(
     rng = random.Random(config.seed)
 
     # Phase 1: 分配 slot + 创建 run 记录
-    slots = _assign_slots(video_ids, task_types, config.per_type, config.family_ratios, rng)
+    slots = _assign_slots(video_ids, task_types, config.per_type)
     logger.info(
         "管线启动: {} slots, {} 视频, {} 任务类型", len(slots), len(video_ids), len(task_types)
     )