Harden saves: retry on commit conflict so answers are never dropped

app.py

@@ -496,6 +496,9 @@ def _reset_save_baseline(participant_file: str, count: int) -> None:
         entry["saved_count"] = count
 
 
+_SAVE_MAX_RETRIES = 6
+
+
 def _save_results(participant_file: str, results: list[dict]) -> None:
     if not HF_TOKEN or not results:
         return
@@ -507,21 +510,34 @@ def _save_results(participant_file: str, results: list[dict]) -> None:
         if len(snapshot) <= entry["saved_count"]:
             return
         frame = pd.DataFrame(snapshot, columns=RESULTS_COLUMNS)
-        buf = io.BytesIO()
-        frame.to_csv(buf, index=False)
-        buf.seek(0)
-        try:
-            api.upload_file(
-                path_or_fileobj=buf,
-                path_in_repo=participant_file,
-                repo_id=RESULTS_REPO,
-                repo_type="dataset",
-                commit_message=f"Update {participant_file} (n={len(snapshot)})",
-            )
-        except Exception as exc:  # noqa: BLE001
-            print(f"[save] WARNING: upload failed for {participant_file} ({exc}).")
-            return
-        entry["saved_count"] = len(snapshot)
+        csv_bytes = frame.to_csv(index=False).encode()
+
+        # Different participants commit to the same repo concurrently, so an
+        # individual upload can still be rejected with a revision conflict.
+        # Retry with backoff so no answer is silently dropped (this was the
+        # original data-loss bug: conflicts were swallowed and never retried).
+        for attempt in range(_SAVE_MAX_RETRIES):
+            try:
+                api.upload_file(
+                    path_or_fileobj=io.BytesIO(csv_bytes),
+                    path_in_repo=participant_file,
+                    repo_id=RESULTS_REPO,
+                    repo_type="dataset",
+                    commit_message=f"Update {participant_file} (n={len(snapshot)})",
+                )
+                entry["saved_count"] = len(snapshot)
+                return
+            except Exception as exc:  # noqa: BLE001
+                wait = 0.5 * (2**attempt) + random.uniform(0, 0.4)
+                print(
+                    f"[save] upload attempt {attempt + 1}/{_SAVE_MAX_RETRIES} "
+                    f"failed for {participant_file} ({exc}); retrying in {wait:.1f}s."
+                )
+                time.sleep(wait)
+        print(
+            f"[save] ERROR: gave up saving {participant_file} after "
+            f"{_SAVE_MAX_RETRIES} attempts (n={len(snapshot)})."
+        )
 
 
 # ---------------------------------------------------------------------------