Add jargon-sensitive readability (Dale-Chall, jargon rate, word familiarity); fix tooltip source term to match the edit

engine.py

@@ -56,6 +56,30 @@ def _read_key(*candidates: Path) -> str:
     return ""
 
 
+# Lexical-familiarity metrics: unlike FKGL (syllable-based, blind to jargon),
+# these reward replacing rare technical words with common everyday ones.
+def _lexical(text: str) -> Dict[str, Any]:
+    out: Dict[str, Any] = {"dale_chall": None, "jargon_rate": None, "zipf_mean": None}
+    t = (text or "").strip()
+    if not t:
+        return out
+    try:
+        import textstat as _ts
+        out["dale_chall"] = round(_ts.dale_chall_readability_score(t), 2)
+    except Exception:
+        pass
+    try:
+        from wordfreq import zipf_frequency as _zf
+        words = re.findall(r"[A-Za-z]+", t.lower())
+        if words:
+            zs = [_zf(w, "en") for w in words]
+            out["zipf_mean"] = round(sum(zs) / len(zs), 2)
+            out["jargon_rate"] = round(sum(1 for z in zs if z < 3.0) / len(words), 4)
+    except Exception:
+        pass
+    return out
+
+
 # --------------------------------------------------------------------------
 # Configuration knobs (overridable via environment before first import).
 # Mirrors run_local.py's CONFIG block.
@@ -580,6 +604,8 @@ class Engine:
                     m.update(cs)
             except Exception:
                 pass
+            # lexical familiarity (captures jargon removal FKGL is blind to)
+            m.update(_lexical(sys_out))
             # reference-based edit/replacement F1 (benchmark mode only)
             if refs:
                 try:
@@ -748,6 +774,13 @@ def _segment_tagged(tagged: str, rat_by_key: Dict, op_by_key: Dict) -> List[Dict
             seg["attribution_weight"] = gt.get("top_attribution_weight")
             seg["src_features"] = gt.get("src_features")
             seg["tgt_features"] = gt.get("tgt_features")
+            # The rationale's source_label is "source (term)" from pick_top_source
+            # (the edit's ACTUAL glossary source) — use its term, not ContextCite's
+            # noisy top-ranked one, so the tooltip source matches the edit.
+            _lbl = gt.get("source_label") or ""
+            _mt = re.search(r"\((.*)\)\s*$", _lbl)
+            if _mt:
+                seg["source_term"] = _mt.group(1).strip()
         if op:
             cc = op.get("contextcite") or {}
             seg["cc_lds"] = cc.get("lds")
@@ -763,11 +796,12 @@ def _segment_tagged(tagged: str, rat_by_key: Dict, op_by_key: Dict) -> List[Dict
                     "passage": (ts.get("passage") or ts.get("text") or "")[:200],
                 })
             seg["ranked_sources"] = ranked
+            # Only fall back to ContextCite's source if the rationale gave none.
             if "source_label" not in seg and ranked:
                 seg["source_short"] = ranked[0]["source"]
                 seg["source_passage"] = ranked[0]["passage"]
-            if ranked and not seg.get("source_term"):
-                seg["source_term"] = ranked[0]["term"]
+                if not seg.get("source_term"):
+                    seg["source_term"] = ranked[0]["term"]
         segments.append(seg)
         last = m.end()
     if last < len(tagged):

static/app.js

@@ -300,6 +300,9 @@
     { key: "fre", label: "Reading ease (FRE)", desc: "0-100 scale", aim: "higher is easier, aim above 60", better: "higher", fmt: n1 },
     { key: "smog", label: "SMOG index", desc: "health-literacy grade", aim: "lower is better", better: "lower", fmt: n1 },
     { key: "coleman_liau", label: "Coleman-Liau", desc: "character-based grade", aim: "lower is better", better: "lower", fmt: n1 },
+    { key: "dale_chall", label: "Dale-Chall grade", desc: "word-familiarity readability (jargon-sensitive)", aim: "lower is better, aim for grade 6", better: "lower", fmt: n1 },
+    { key: "jargon_rate", label: "Jargon words", desc: "share of rare/technical words", aim: "lower is better, 0 = none left", better: "lower", fmt: pct, bar01: true },
+    { key: "zipf_mean", label: "Word familiarity", desc: "average word commonness (Zipf)", aim: "higher is better, everyday words", better: "higher", fmt: n1 },
     { key: "mepr", label: "Entity preservation", desc: "medical facts kept (MEPR)", aim: "higher is better, keep the facts", better: "higher", fmt: pct, bar01: true },
     { key: "nli_faith", label: "NLI faithfulness", desc: "no hallucination vs input", aim: "higher is better, no made-up facts", better: "higher", fmt: pct, bar01: true },
     { key: "critical_error", label: "Critical safety error", desc: "lost negation / number / side", aim: "lower is better, 0 is safe", better: "lower", fmt: pct },