feat: check de colisión de tema (topic.collision) en rules + validate

Dos posts sobre el mismo caso se canibalizan en la SERP (2026-07-10: se
publicó un segundo Kecksburg con otro ya programado, y había un título
casi gemelo del de Malmstrom en la cola). Nueva función pura
topic_collision(post, corpus) en seo_rules.py — corpus-aware, fuera de
RULES — que dispara por: caso+año compartidos (años < 2020; los de la
era de noticias no son identidad de caso), hooks de título ≥70%
similares (calibrado: el par nuclear da 0.742, el par legítimo más
cercano 0.65), o slugs ≥50% solapados. seo_validate.py obtiene el corpus
via ghst-en (--limit all) y lo reporta como grupo topic_collision;
--no-collision lo omite para follow-ups intencionados.

Calibrado contra los 36 posts reales: dispara solo el duplicado
Kecksburg (HIGH) y el par Grusch preview/evento (MED, intencionado).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

seo_rules.py

@@ -178,6 +178,112 @@ def r_jsonld(p):
     return [Violation("jsonld.missing", MED, "no BlogPosting JSON-LD", "add JSON-LD")]
 
 
+# ---- topic collision (corpus-aware; NOT in RULES) ---------------------------
+# Two posts about the same case cannibalize each other in the SERP (2026-07-10:
+# a second Kecksburg post was published while another sat scheduled; a "When
+# Nuclear ... Went/Go Silent" near-twin title was already queued). RULES functions
+# are (post) -> violations; this one also needs the rest of the site, so callers
+# (seo_validate.py) pass the corpus explicitly: published + scheduled posts as
+# dicts with at least {id, title, slug}.
+
+TOPIC_STOPWORDS = {
+    # english glue
+    "the", "a", "an", "of", "and", "in", "at", "on", "to", "that", "what",
+    "when", "who", "why", "how", "its", "his", "her", "their", "our", "one",
+    "still", "cant", "couldnt", "went", "go", "goes", "most", "from", "with",
+    "they", "them", "these", "this", "are", "were", "was", "is", "be", "been",
+    "has", "have", "had", "but", "for", "all", "than", "then", "ever", "never",
+    # domain-generic (present in half the catalog — carry no case identity)
+    "ufo", "ufos", "uap", "uaps", "incident", "incidents", "case", "cases",
+    "file", "files", "mystery", "declassified", "declassification", "pentagon",
+    "government", "military", "congress", "secret", "program", "investigation",
+    "evidence", "witness", "witnesses", "document", "documents", "documented",
+    "unexplained", "encounter", "sighting", "sightings", "alien", "aliens",
+    "phenomena", "aerial", "unidentified", "extraordinary", "americas",
+    "american", "video", "footage",
+}
+# 0.70 calibrated 2026-07-10: the "When Nuclear Weapons Go / Arsenal Went
+# Silent" near-twin pair scores 0.742 (char-level penalizes weapons/arsenal);
+# the closest legit-distinct pair in the catalog scores 0.65.
+TITLE_HOOK_SIM_MIN = 0.70   # SequenceMatcher on the pre-colon hook
+SLUG_JACCARD_MIN   = 0.5    # shared slug-token ratio
+# Years >= this are "news era", not case identity: every contemporary post
+# carries the current year (PURSUE 2026, Grusch 2026...) without being the same
+# story. Case years in the catalog run 1947-2019.
+NEWS_YEAR_MIN = 2020
+
+_YEAR_RE = re.compile(r"\b(19|20)\d{2}\b")
+
+
+def _tokens(text):
+    return set(re.findall(r"[a-z0-9]+", _s(text).lower()))
+
+
+def _case_years(title, slug):
+    """Historical case years (pre news-era) found in title+slug."""
+    return {m.group(0) for m in _YEAR_RE.finditer(title + " " + slug)
+            if int(m.group(0)) < NEWS_YEAR_MIN}
+
+
+def _sig_tokens(title, slug):
+    """Case-identity tokens: title+slug minus glue/domain words, years and
+    fragments shorter than 3 chars (possessive 's', initials...)."""
+    toks = _tokens(title) | _tokens(slug.replace("-", " "))
+    return {t for t in toks
+            if len(t) >= 3 and t not in TOPIC_STOPWORDS and not _YEAR_RE.fullmatch(t)}
+
+
+def _hook(title):
+    return _s(title).split(":")[0].strip().lower()
+
+
+def topic_collision(post, corpus):
+    """Compare one candidate post against the site corpus → list[Violation].
+
+    Fires when the candidate and an existing post look like the same story:
+      - share a case year AND a case-identity token (Kecksburg+1965), or
+      - their pre-colon title hooks read nearly the same, or
+      - their slugs share most of their tokens.
+    """
+    from difflib import SequenceMatcher
+
+    out = []
+    c_years = _case_years(_s(post.get("title")), _s(post.get("slug")))
+    c_sig = _sig_tokens(post.get("title"), _s(post.get("slug")))
+    c_hook = _hook(post.get("title"))
+    c_slug_toks = _tokens(_s(post.get("slug")).replace("-", " "))
+
+    for other in corpus:
+        if other.get("id") == post.get("id"):
+            continue
+        o_title, o_slug = _s(other.get("title")), _s(other.get("slug"))
+        o_years = _case_years(o_title, o_slug)
+        o_sig = _sig_tokens(o_title, o_slug)
+
+        reasons = []
+        if (c_years & o_years) and (c_sig & o_sig):
+            shared = ", ".join(sorted(c_sig & o_sig)[:3] + sorted(c_years & o_years))
+            reasons.append((HIGH, f"same case + year ({shared})"))
+        hook_sim = SequenceMatcher(None, c_hook, _hook(o_title)).ratio()
+        if c_hook and hook_sim >= TITLE_HOOK_SIM_MIN:
+            reasons.append((MED, f"title hooks {hook_sim:.0%} similar"))
+        o_slug_toks = _tokens(o_slug.replace("-", " "))
+        union = c_slug_toks | o_slug_toks
+        if union:
+            jac = len(c_slug_toks & o_slug_toks) / len(union)
+            if jac >= SLUG_JACCARD_MIN:
+                reasons.append((MED, f"slugs {jac:.0%} overlapping"))
+
+        if reasons:
+            sev = max(s for s, _ in reasons)
+            why = "; ".join(r for _, r in reasons)
+            out.append(Violation(
+                "topic.collision", sev,
+                f"collides with [{other.get('status', '?')}] \"{o_title[:60]}\" — {why}",
+                "merge, retitle to a distinct angle, or interlink deliberately"))
+    return out
+
+
 RULES = [
     r_meta_title,
     r_meta_description,