A8: 文案按A/B/C三套正交叙事生成,避免套路化重复

- constants: 新增 TEXT_NARRATIVE_BY_STRATEGY(A痛点/B场景/C成分),与图片侧同轴
- build_prompt: 加 strategy_narrative 参数并注入 prompt
- text_variants: 全链路透传(含优化轮)
- run_text_generation: 改循环三套,text_count均摊(divmod余前补),跨套去重,打_strategy标记
- TextCandidate: 加 strategy String(4) 字段 + 迁移021(已upgrade head)
- packaging: 打包按strategy精准配对文图(texts_by_strategy映射+三层兜底)
- SSE text_candidate 事件携带 strategy

独立agent交叉验证7改造点全过,边界(text_count<3/无别名/不截断)无must-fix

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/app/workers/pipeline_io.py

@@ -46,7 +46,7 @@ def run_text_generation(db, clients, task, product_dict: dict, flywheel_fragment
     from app.models.flywheel import AiCallLog
     from app.models.workspace import UserApiKey
     from app.constants.enums import CandidateSource, BannedWordStatus
-    from app.services.ai_engine.constants import QUALITY_PASS_SCORE
+    from app.services.ai_engine.constants import QUALITY_PASS_SCORE, TEXT_NARRATIVE_BY_STRATEGY
 
     banned_rows = db.query(BannedWord).filter(
         BannedWord.workspace_id == workspace_id
@@ -61,20 +61,37 @@ def run_text_generation(db, clients, task, product_dict: dict, flywheel_fragment
     ).first()
     key_id = key_row.id if key_row else None
 
+    # A8：文案按 A/B/C 三套正交叙事分别生成，每套不同角度避免套路化重复，
+    # 与图片侧同轴(A痛点/B场景/C成分)。text_count 均摊三套(余数前补)，
+    # 逐套把已生成的喂作 previous_copies 做跨套去重。
+    _strategies = ("A", "B", "C")
+    _base, _rem = divmod(task.text_count, 3)
+    _per = {s: _base + (1 if i < _rem else 0) for i, s in enumerate(_strategies)}
+
     t0 = time.monotonic()
     llm_success = True
-    try:
-        candidates_raw = asyncio.run(generate_text_variants(
-            llm_client=clients,
-            product=product_dict,
-            count=task.text_count,
-            banned_word_rows=banned_dicts,
-            flywheel_context=flywheel_fragment,
-        ))
-    except Exception as exc:
-        llm_success = False
-        logger.error("generate_text_variants 失败: %s", exc)
-        candidates_raw = []
+    candidates_raw: list = []
+    for s in _strategies:
+        n = _per[s]
+        if n <= 0:
+            continue
+        try:
+            part = asyncio.run(generate_text_variants(
+                llm_client=clients,
+                product=product_dict,
+                count=n,
+                previous_copies=candidates_raw,
+                banned_word_rows=banned_dicts,
+                flywheel_context=flywheel_fragment,
+                strategy_narrative=TEXT_NARRATIVE_BY_STRATEGY.get(s, ""),
+            ))
+        except Exception as exc:
+            llm_success = False
+            logger.error("generate_text_variants(套%s) 失败: %s", s, exc)
+            part = []
+        for c in part:
+            c["_strategy"] = s
+        candidates_raw.extend(part)
     latency_ms = int((time.monotonic() - t0) * 1000)
 
     # 写 ai_call_logs（留痕，不含明文key）
@@ -113,6 +130,7 @@ def run_text_generation(db, clients, task, product_dict: dict, flywheel_fragment
             workspace_id=workspace_id,
             task_id=task.id,
             source=CandidateSource.AI,
+            strategy=c.get("_strategy"),
             angle_label=c.get("angle_label") or c.get("angle", ""),
             content=json.dumps(c, ensure_ascii=False),
             score_json=json.dumps(c.get("score_detail", []), ensure_ascii=False),
@@ -124,6 +142,7 @@ def run_text_generation(db, clients, task, product_dict: dict, flywheel_fragment
         seq += 1
         push_fn(task.id, workspace_id, "text_candidate", {
             "candidate_id": tc.id, "angle_label": tc.angle_label,
+            "strategy": tc.strategy,
             "content": c.get("content", ""), "score": score,
         }, seq)
         seq += 1