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_validate.py

@@ -39,9 +39,31 @@ GROUPS = [
     ("feature_image",     ("feature_image.",)),
     ("feature_image_alt", ("feature_image_alt.",)),
     ("internal_links",    ("internal_links.",)),
+    ("topic_collision",   ("topic.",)),
 ]
 
 
+def fetch_corpus():
+    """All published+scheduled posts (id, title, slug, status) via ghst-en, for
+    the topic-collision check. Returns [] (with a warning) if the fetch fails —
+    the validator still runs the per-post rules."""
+    fd, path = tempfile.mkstemp(suffix=".json", prefix="seo_corpus_")
+    os.close(fd)
+    try:
+        cmd = (f"ghst-en --json post list --limit all "
+               f"--fields id,title,slug,status > {path}")
+        r = subprocess.run(cmd, shell=True, stderr=subprocess.PIPE, text=True)
+        if r.returncode != 0:
+            print(f"  [warn] corpus fetch failed — topic collision NOT checked:\n{r.stderr}",
+                  file=sys.stderr)
+            return []
+        data = json.load(open(path))
+        posts = data.get("posts", data) if isinstance(data, dict) else data
+        return [p for p in posts if p.get("status") in ("published", "scheduled")]
+    finally:
+        os.unlink(path)
+
+
 def fetch_one(selector, value):
     """Fetch a single post (draft or published) via ghst-en → post dict.
 
@@ -76,10 +98,13 @@ def load_json(path):
     return _unwrap(json.load(open(path)))
 
 
-def validate(post):
-    """Run the shared engine and print a per-group PASS/FAIL report.
+def validate(post, corpus=None):
+    """Run the shared engine (plus the corpus-aware topic-collision check when a
+    corpus is given) and print a per-group PASS/FAIL report.
     Returns the number of actionable (non-INFO) violations."""
     violations = R.check_post(post)
+    if corpus:
+        violations.extend(R.topic_collision(post, corpus))
     actionable = [v for v in violations if v.severity > R.INFO]
     info = [v for v in violations if v.severity == R.INFO]
 
@@ -129,6 +154,8 @@ def main():
     g.add_argument("--slug")
     g.add_argument("--id")
     g.add_argument("--json", metavar="FILE")
+    ap.add_argument("--no-collision", action="store_true",
+                    help="skip the topic-collision check (intentional follow-up piece)")
     args = ap.parse_args()
 
     if args.json:
@@ -138,7 +165,8 @@ def main():
     else:
         post = fetch_one("id", args.id)
 
-    n = validate(post)
+    corpus = [] if args.no_collision else fetch_corpus()
+    n = validate(post, corpus)
     sys.exit(1 if n else 0)