fix(goldset): score panel open by default + sparse-negatives hint

The validator score panel was collapsed by default, so taggers thought nothing
was happening. Now open by default, with a caption explaining the metrics
measure "not-a-holding" detection and become meaningful as more "לא הלכה" items
are tagged (showing the current negative count while it's small).

Verified: tsc --noEmit 0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

web-ui/src/components/goldset/goldset-panel.tsx

@@ -71,9 +71,12 @@ function isTagged(it: GoldsetItem): boolean {
 
 function ScorePanel({ batch }: { batch: string }) {
   const { data } = useGoldsetScore(batch);
-  const [open, setOpen] = useState(false);
+  const [open, setOpen] = useState(true);
   if (!data || data.labeled === 0) return null;
   const rows = Object.entries(data.validators);
+  // negatives so far (truly "not a holding") = tp+fn of any validator.
+  const af = data.validators.any_flag;
+  const negatives = af ? af.tp + af.fn : 0;
   return (
     <div className="rounded-lg border border-rule bg-surface">
       <button
@@ -88,6 +91,12 @@ function ScorePanel({ batch }: { batch: string }) {
       </button>
       {open && (
         <div className="px-4 pb-3 overflow-x-auto">
+          <p className="text-[0.72rem] text-ink-muted mb-2">
+            המדדים מודדים זיהוי <strong>"לא-הלכה"</strong> (יישום / ציטוט-קטוע / אי-הכרעה...).
+            {negatives < 10
+              ? ` עד כה תויגו רק ${negatives} פריטי "לא הלכה" — המספרים יהפכו משמעותיים ככל שיצטברו עוד (במיוחד מבקט המסומנים).`
+              : " precision גבוה = מעט אזעקות-שווא; recall גבוה = תופס את רוב ה'לא-הלכה'."}
+          </p>
           <table className="w-full text-sm tabular-nums">
             <thead>
               <tr className="text-ink-muted text-[0.72rem] border-b border-rule">