feat(classify): tighten scoring with reaction co-occurrence and date negatives

强正向 reaction 需标题同时含 require_title；移除听/看等宽泛词，新增弱正向组；charge 命中>=2 记+3、==1 记+1；负向补充录播/录像并新增标题带日期强负向；时长仅在已命中 reaction/charge 时加权；discover 关键词改为单个咻咻满、pages=3、window_days=7，notify.min_score=4；同步示例配置、project.md/README 与相关测试。

reaction_board/classify.py

@@ -47,11 +47,16 @@ def parse_duration(value: Any) -> int:
     return 0
 
 
-def contains_any(title: str, words: Any) -> bool:
+def matched_words(title: str, words: Any) -> list[str]:
+    """返回标题命中（归一化子串）的词，保持 ``words`` 顺序、跳过空词。"""
     normalized = normalize_title(title)
     if not isinstance(words, (list, tuple)):
-        return False
-    return any(normalize_title(str(word)) in normalized for word in words if word)
+        return []
+    return [str(word) for word in words if word and normalize_title(str(word)) in normalized]
+
+
+def contains_any(title: str, words: Any) -> bool:
+    return bool(matched_words(title, words))
 
 
 def passes_required_title(title: str, require_title: str) -> bool:
@@ -60,6 +65,30 @@ def passes_required_title(title: str, require_title: str) -> bool:
     return normalize_title(require_title) in normalize_title(title)
 
 
+# 标题带日期通常意味着录播/回放/录像（按具体某场直播命名），作为强负向信号。
+# 覆盖 `2026-10-28` / `2026/10/25` / `2026.10.26` / `2024年11月24日` / `10月25日`。
+_DATE_PATTERNS = (
+    r"\d{4}\s*[-/.]\s*\d{1,2}\s*[-/.]\s*\d{1,2}",
+    r"\d{4}\s*年\s*\d{1,2}\s*月(?:\s*\d{1,2}\s*日)?",
+    r"\d{1,2}\s*月\s*\d{1,2}\s*日",
+)
+_DATE_RE = re.compile("|".join(_DATE_PATTERNS))
+
+# 各信号的权重（见 project.md §3.3）。
+REACTION_SCORE = 3
+CHARGE_MULTI_SCORE = 3
+CHARGE_SINGLE_SCORE = 1
+WEAK_SCORE = 1
+NEGATIVE_SCORE = 6
+DATE_SCORE = 6
+SELF_SCORE = 3
+
+
+def date_hits(title: str) -> list[str]:
+    """返回标题中的日期片段，用于判定「带日期≈录播/回放」。"""
+    return _DATE_RE.findall(title or "")
+
+
 def score_item(
     *,
     title: str,
@@ -67,23 +96,43 @@ def score_item(
     duration_seconds: int,
     classify: Mapping[str, Any],
 ) -> int:
-    """按 project.md §3.3 计算候选分数。"""
+    """按 project.md §3.3 计算候选分数。
+
+    强正向 ``reaction`` 只在标题同时命中 ``require_title`` 时成立；时长只作为
+    已存在 reaction / charge 信号时的加权，避免纯长录像仅靠时长进池。
+    """
     score = 0
-    if contains_any(title, classify.get("positive")):
-        score += 2
-    if contains_any(title, classify.get("charge")):
-        score += 2
+    require_title = str(classify.get("require_title") or "")
 
-    boost = classify.get("duration_boost") or {}
-    gt_60 = int(boost.get("gt_60min", 2))
-    gt_30 = int(boost.get("gt_30min", 1))
-    if duration_seconds > 60 * 60:
-        score += gt_60
-    elif duration_seconds > 30 * 60:
-        score += gt_30
+    reaction_matched = bool(matched_words(title, classify.get("reaction")))
+    reaction_hit = reaction_matched and (
+        not require_title or passes_required_title(title, require_title)
+    )
+    if reaction_hit:
+        score += REACTION_SCORE
 
-    if contains_any(title, classify.get("negative")):
-        score -= 5
+    charge_hits = matched_words(title, classify.get("charge"))
+    if len(charge_hits) >= 2:
+        score += CHARGE_MULTI_SCORE
+    elif len(charge_hits) == 1:
+        score += CHARGE_SINGLE_SCORE
+
+    if matched_words(title, classify.get("weak_positive")):
+        score += WEAK_SCORE
+
+    if reaction_hit or charge_hits:
+        boost = classify.get("duration_boost") or {}
+        gt_60 = int(boost.get("gt_60min", 2))
+        gt_30 = int(boost.get("gt_30min", 1))
+        if duration_seconds > 60 * 60:
+            score += gt_60
+        elif duration_seconds > 30 * 60:
+            score += gt_30
+
+    if matched_words(title, classify.get("negative")):
+        score -= NEGATIVE_SCORE
+    if classify.get("date_negative", True) and date_hits(title):
+        score -= DATE_SCORE
     if up_mid is not None and int(up_mid) == int(classify.get("self_mid", 37754047)):
-        score -= 3
+        score -= SELF_SCORE
     return score

reaction_board/config.py

@@ -25,12 +25,11 @@ from .client_ip import (
 DEFAULT_CLASSIFY: dict[str, Any] = {
     "require_title": "咻咻满",
     "self_mid": 37754047,
-    "positive": [
-        "reaction", "re", "反应", "看", "听", "初听", "首次",
-        "点评", "锐评", "路人", "震惊", "双人",
-    ],
-    "charge": ["充电", "加急", "专场"],
-    "negative": ["歌切", "切片", "纯享", "搬运", "完整版"],
+    "reaction": ["reaction", "反应"],
+    "charge": ["加急", "特急", "充电", "专场", "转场"],
+    "weak_positive": ["双人", "路人", "第一次", "首次"],
+    "negative": ["歌切", "切片", "纯享", "搬运", "完整版", "录播", "录像"],
+    "date_negative": True,
     "duration_boost": {"gt_30min": 1, "gt_60min": 2},
 }
 
@@ -87,11 +86,11 @@ DEFAULT_CONFIG: dict[str, Any] = {
         "candidate_ttl_days": 7,
     },
     "discover": {
-        "keywords": ["咻咻满 reaction", "咻咻满 充电"],
-        "pages": 2,
-        "window_days": 14,
+        "keywords": ["咻咻满"],
+        "pages": 3,
+        "window_days": 7,
     },
-    "notify": {"min_score": 2, "max_items_per_message": 10},
+    "notify": {"min_score": 4, "max_items_per_message": 10},
     "classify": DEFAULT_CLASSIFY,
     "heat_weights": DEFAULT_HEAT_WEIGHTS,
     "analytics": DEFAULT_ANALYTICS,
@@ -500,8 +499,12 @@ def _build_config(data: Mapping[str, Any], path: Path) -> AppConfig:
     classify["self_mid"] = _as_int(
         classify.get("self_mid", 37754047), "classify.self_mid", minimum=1
     )
-    for key in ("positive", "charge", "negative"):
+    for key in ("reaction", "charge", "weak_positive", "negative"):
         classify[key] = _as_str_list(classify.get(key, []), f"classify.{key}")
+    date_negative = classify.get("date_negative", True)
+    if not isinstance(date_negative, bool):
+        raise ConfigError("classify.date_negative 必须是布尔值")
+    classify["date_negative"] = date_negative
     boost = classify.get("duration_boost", {})
     if not isinstance(boost, Mapping):
         raise ConfigError("classify.duration_boost 必须是对象")