change: paper results win on conflict (official results immutable)

Per request, flip the leaderboard merge so the bundled paper results are
authoritative: on a (model, adapter) clash the paper row wins, and community
submissions from braindecode/contents can only ADD new model/strategy combos,
never override an official result.

backend/app/services/leaderboard.py

@@ -30,13 +30,14 @@ def _merge_results(
 ) -> List[Dict[str, Any]]:
     """Merge curated/base results with new ones, deduped by (model, adapter).
 
-    ``base`` is the bundled real/paper results; ``overlay`` is new community
-    results from the HF contents dataset. Overlay rows override base rows that
-    share the same (fullname, adapter) so a freshly evaluated result supersedes
-    the baseline, while every other base result is preserved.
+    ``base`` is the bundled real/paper results (authoritative & immutable);
+    ``overlay`` is new community results from the HF contents dataset. The paper
+    results always win on a (fullname, adapter) clash, so community submissions
+    can only ADD new model/strategy combinations — they never override an
+    official result.
     """
     merged: Dict[tuple, Dict[str, Any]] = {}
-    for row in list(base) + list(overlay):  # overlay last => wins on conflict
+    for row in list(overlay) + list(base):  # base last => paper results win on conflict
         merged[(row.get("fullname"), row.get("adapter"))] = row
     return list(merged.values())
 
@@ -48,9 +49,10 @@ class LeaderboardService:
     async def fetch_raw_data(self) -> List[Dict[str, Any]]:
         """Merge the bundled real/paper results with new community results.
 
-        ``sample_results.json`` (the paper results) is the immutable base. New
-        results from the HF ``contents`` dataset are overlaid on top so the board
-        shows BOTH; on a (model, adapter) clash the new result wins. Falls back
+        ``sample_results.json`` (the paper results) is the authoritative, immutable
+        base. New results from the HF ``contents`` dataset are overlaid so the
+        board shows BOTH; on a (model, adapter) clash the PAPER result wins, so
+        community submissions can only add new model/strategy combos. Falls back
         to base-only when the contents dataset isn't available yet.
         """
         try:

backend/tests/test_leaderboard_merge.py

@@ -1,21 +1,21 @@
 from app.services.leaderboard import _merge_results
 
 
-def test_merge_overlay_wins_on_conflict():
+def test_merge_paper_base_wins_on_conflict():
     base = [
         {"fullname": "BIOT", "adapter": "full_finetune", "Average ⬆️": 30.0},
         {"fullname": "LaBraM", "adapter": "probe", "Average ⬆️": 12.0},
     ]
     overlay = [
-        {"fullname": "BIOT", "adapter": "full_finetune", "Average ⬆️": 40.0},  # overrides base
-        {"fullname": "MyModel", "adapter": "lora", "Average ⬆️": 50.0},        # brand new
+        {"fullname": "BIOT", "adapter": "full_finetune", "Average ⬆️": 40.0},  # same combo as a paper row
+        {"fullname": "MyModel", "adapter": "lora", "Average ⬆️": 50.0},        # brand new combo
     ]
     out = _merge_results(base, overlay)
     by = {(r["fullname"], r["adapter"]): r for r in out}
     assert len(out) == 3
-    assert by[("BIOT", "full_finetune")]["Average ⬆️"] == 40.0   # new result wins
+    assert by[("BIOT", "full_finetune")]["Average ⬆️"] == 30.0   # paper result wins (immutable)
     assert by[("LaBraM", "probe")]["Average ⬆️"] == 12.0          # paper result preserved
-    assert ("MyModel", "lora") in by                              # new model added
+    assert ("MyModel", "lora") in by                              # new combo added
 
 
 def test_merge_empty_overlay_returns_base():