hf_raw: fall back to cached blobs when the mirror listing degrades

The tree API intermittently returns truncated listings (9 parquets
listed as [.gitattributes, README.md] at night); previously that failed
the bench even though 5 of the files were already in the shared blob
store. Empty selection now uses matching cached blobs with a notice;
only a truly cold cache still raises.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/data/loader.py

@@ -422,11 +422,32 @@ def _hf_download(repo: str, path: str, dest_dir: Path) -> Path:
 def _load_from_hf_raw(spec: DatasetSpec, raw_dir: Optional[Path] = None) -> List[Dict[str, Any]]:
     import hashlib
 
+    # the mirror's tree API degrades at night and returns TRUNCATED listings
+    # (a repo with 9 parquets listed as [.gitattributes, README.md]); the
+    # blobs from a previous run are already in the shared store -- use them
+    blob_dir = get_cache_root() / '.raw' / hashlib.md5(spec.source.encode()).hexdigest()[:10]
+
+    def _cached_blobs() -> List[str]:
+        if not blob_dir.exists():
+            return []
+        have = [f.name for f in blob_dir.iterdir()
+                if os.path.splitext(f)[1] in _SUPPORTED_EXTS]
+        pref = [f for f in have if f.startswith(spec.subset)] or have
+        return pref
+
     files = _hf_list_files(spec.source)
-    if not files:
-        raise FileNotFoundError(f'no files found on HF dataset {spec.source!r}')
-    selected = _hf_match_files(spec, files)
+    selected = _hf_match_files(spec, files) if files else []
     if not selected:
+        cached = _cached_blobs()
+        if cached:
+            print(f'· listing unavailable/degraded -- using {len(cached)} '
+                  f'cached file(s) from {blob_dir}', flush=True)
+            records: List[Dict[str, Any]] = []
+            for f in sorted(cached):
+                records.extend(_read_file(str(blob_dir / f)))
+            if raw_dir is not None:
+                _link_or_copy_all([blob_dir / f for f in sorted(cached)], raw_dir)
+            return records
         raise FileNotFoundError(
             f'no data file for subset={spec.subset!r} split={spec.split!r} in '
             f'HF {spec.source!r}. Available (first 10): {files[:10]}'