feat: F3.4 semantic_collapse — n-gram Jaccard dedup defeats synonym padding

swarmmetrics.py

@@ -522,6 +522,51 @@ if __name__ == "__main__":
               f"sent={ns.total_sent}, recv={ns.total_received}{shadow}")
 
 
+# ── Semantic collapse (F3.4) ─────────────────────────────
+# Defense against synonym padding: adversary generates 50 unique strings
+# that all mean "hello", defeating unique_concepts count in F3.3.
+# Collapse semantically similar concepts before counting.
+# Atomic Raven (Colony, 2026-07-28) identified the attack vector.
+
+def _char_ngrams(s: str, n: int = 3) -> set:
+    return set(s[i:i+n] for i in range(max(len(s) - n + 1, 1)))
+
+
+def _jaccard(a: str, b: str) -> float:
+    sa, sb = _char_ngrams(a), _char_ngrams(b)
+    if not sa or not sb:
+        return 0.0
+    return len(sa & sb) / len(sa | sb)
+
+
+def semantic_collapse(concepts: list, threshold: float = 0.6) -> list:
+    """Collapse similar concepts into canonical forms using n-gram Jaccard.
+
+    A cheap proxy for semantic similarity. Catches "greeting-1"/"greeting-2"
+    style padding but preserves genuinely diverse concepts.
+
+    threshold: Jaccard similarity above which concepts merge.
+    0.6 collapses trivial suffix variations, keeps distinct concepts.
+
+    Returns deduplicated concept list (canonical forms only).
+    """
+    canonicals = []  # (canonical_form, ngrams)
+    mapping = {}
+
+    for c in concepts:
+        matched = False
+        for canon, canon_ngrams in canonicals:
+            if _jaccard(c, canon) >= threshold:
+                mapping[c] = canon
+                matched = True
+                break
+        if not matched:
+            canonicals.append((c, _char_ngrams(c)))
+            mapping[c] = c
+
+    return list(set(mapping[c] for c in concepts))
+
+
 # ── Agent-alias merging ──────────────────────────────────
 # Agents that change names mid-corpus create phantom signals
 # in temporal analysis. merge_aliases() canonicalizes from_id/to_id