悄悄话留言自查：孕周和内容对齐、去掉过来人身份矛盾的回复、名字不撞套路

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tools/whcm/build.py

@@ -14,22 +14,56 @@ def wk(p):
 # 内容多是胎动、产检、待产（孕中晚期的事）→ 主评论只用孕 24 周以上的人设，免得「孕 6 周」说「头朝下了」
 preg = sorted([i for i, p in enumerate(d['p']) if wk(p) >= 24], key=lambda i: (used[i], i))
 anyp = sorted(range(len(d['p'])), key=lambda i: (used[i], i))
+# 11-07 用户看评论：名字套路太像（4 个「在挑名字」、4 个「今天也开心」、3 个「水獭」、虎虎的老母亲/虎虎妈妈…）→ 同一帖里前两个字、后两个字都不撞
+SIG = set()
+def fresh(i):
+    n = d['p'][i][0]; ks = {'h:' + n[:2], 't:' + n[-2:]}
+    return not (ks & SIG)
+def take(pool, k):
+    out = []
+    for i in pool:
+        if len(out) >= k: break
+        if fresh(i):
+            n = d['p'][i][0]; SIG.update({'h:' + n[:2], 't:' + n[-2:]}); out.append(i)
+    for i in pool:   # 不够就只要求后两个字不撞
+        if len(out) >= k: break
+        n = d['p'][i][0]
+        if i not in out and 't:' + n[-2:] not in SIG: SIG.add('t:' + n[-2:]); out.append(i)
+    for i in pool:   # 还不够就随它
+        if len(out) >= k: break
+        if i not in out: out.append(i)
+    return out
+preg = take(preg, 160)
 mains = preg[:200]; rnd.shuffle(mains)
-reps = anyp[:300]; rnd.shuffle(reps)
+anyp = [i for i in anyp if i not in set(preg) and not re.match(r'准爸爸', d['p'][i][2])]   # 回复里不放「准爸爸」（还没娃，说不了打针）
+reps = take(anyp, 31); rnd.shuffle(reps)
 lines = [l.strip() for l in open('tools/whcm/lines.txt', encoding='utf-8') if l.strip()]
 assert len(lines) == 100, len(lines)
 out, ri = [], 0
+# 内容要和孕周对得上：按关键词给每条定孕周范围，从打乱后的人设里挑第一个落在范围内的
+RULES = [(r'糖耐', 24, 28), (r'预产期在月底', 37, 39), (r'待产包|入盆|无痛|胎位|头朝下|小床|产假|月子餐|育儿嫂|报纸|满月|儿科医生|插座|剪了标签|绘本', 32, 39),
+         (r'踢的位置越来越低|孕妇照|羊水|吸奶器|数胎动|数你|产褥垫|剪指甲|洗了第一批衣服|婴儿车', 28, 39), (r'四维', 24, 30)]
+def rng(t):
+    for pat, lo, hi in RULES:
+        if re.search(pat, t): return lo, hi
+    return 24, 39
+left = list(mains)
+def pick(t):
+    lo, hi = rng(t)
+    for i in left:
+        if lo <= wk(d['p'][i]) <= hi: left.remove(i); return i
+    raise SystemExit('没有合适人设：' + t)
 for k, l in enumerate(lines):
     parts = l.split('||')
     likes = max(3, int(520 * (0.965 ** k)) + rnd.randint(-18, 18))
     rr = []
     for t in parts[1:]:
         rr.append([reps[ri], t]); ri += 1
-    out.append([mains[k], parts[0], likes, rr])
+    out.append([pick(parts[0]), parts[0], likes, rr])
 # 11-07 用户：「评论里也加一些孕早期的孕妈，看到别人有胎动也有期待，也有自己的话对孩子说；都要真人感」
 #  early.txt = 孕 6～13 周人设的留言（回复用孕中晚期/宝妈人设，像过来人）；均匀插进前 100 条中间，不全堆在后面
 #  early_rep.txt = 「原留言开头|早孕妈妈的回复」，挂在已有留言楼里（「看你们说胎动好羡慕」这类）
-early_p = sorted([i for i, p in enumerate(d['p']) if 6 <= wk(p) <= 13 and i not in set(reps)], key=lambda i: (used[i], i))
+early_p = take(sorted([i for i, p in enumerate(d['p']) if 6 <= wk(p) <= 13 and i not in set(reps) and not re.search(r'待产|宝妈|老母亲|养娃|学抱娃|拼尿布|麻麻|大宝', p[0])], key=lambda i: (used[i], i)), 51)
 er = random.Random(20261108); er.shuffle(early_p)
 ei = 0
 for t in [l.strip() for l in open('tools/whcm/early_rep.txt', encoding='utf-8') if l.strip()]: