all the changes

cmd/needle-sidecar/eval/run_eval.py

@@ -4,8 +4,7 @@
 Usage: run_eval.py [--url http://127.0.0.1:5005] [--golden eval/golden.jsonl] [--json out.json]
 
 Each golden row: {"id", "expected", "text"}; `expected` is a sidecar action name
-or "none" when no tool should fit. The sidecar has no fallback tool, so "none"
-rows measure how often it force-fits (see the confidence column).
+including "needs_review" for the fallback.
 """
 import argparse, json, os, sys, time, urllib.request
 from collections import Counter, defaultdict
@@ -29,7 +28,7 @@ def main():
     ap.add_argument("--golden", default=os.path.join(HERE, "golden.jsonl"))
     ap.add_argument("--json", help="write per-row results here")
     ap.add_argument("--min-conf", type=float, default=0.0,
-                    help="predictions below this confidence count as 'none'")
+                    help="predictions below this confidence count as needs_review")
     a = ap.parse_args()
 
     rows = [json.loads(l) for l in open(a.golden) if l.strip()]
@@ -41,10 +40,10 @@ def main():
         except Exception as e:
             print(f"sidecar error on {row['id']}: {e}", file=sys.stderr)
             sys.exit(2)
-        pred = out.get("action", "unclassified")
+        pred = out.get("action", "needs_review")
         conf = out.get("confidence", 0.0)
-        if conf < a.min_conf or pred == "unclassified":
-            pred = "none"
+        if conf < a.min_conf:
+            pred = "needs_review"
         lat.append(out.get("latency_ms", 0.0))
         confusion[(row["expected"], pred)] += 1
         results.append({**row, "predicted": pred, "confidence": conf, "ok": pred == row["expected"]})