perf: parallelize --repair with a thread pool (--workers, default 8)

Each repaired file is an independent yt-dlp metadata round-trip, so repair is
network-bound; run them concurrently via ThreadPoolExecutor. Adds --workers
(default 8) to cap concurrency and a progress line every 100 files. At ~50k
tracks this turns a ~day-long sequential run into hours. Lower --workers if
YouTube rate-limits (429/403).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_repair.py

@@ -336,3 +336,34 @@ def test_repair_file_dry_run_does_not_rename(tmp_path, monkeypatch):
 
 def test_fs_safe_replaces_slash():
     assert "/" not in mf._fs_safe("AC/DC Live")
+
+
+# ---- parallel repair ----
+def test_repair_library_parallel_visits_all(tmp_path, monkeypatch):
+    root = tmp_path
+    n = 50
+    for i in range(n):
+        d = root / f"Artist{i}" / "youtube"
+        d.mkdir(parents=True)
+        (d / f"T{i} [{YT_ID}].opus").write_text("x")
+
+    import threading
+    seen = set()
+    lock = threading.Lock()
+
+    def fake(path, source, dry_run):
+        with lock:
+            seen.add(path)
+        return ["album=X"]
+    monkeypatch.setattr(mf, "repair_file", fake)
+    scanned, changed = mf.repair_library(str(root), dry_run=False, workers=8)
+    assert scanned == n and changed == n
+    assert len(seen) == n
+
+
+def test_repair_library_default_workers_still_works(tmp_path, monkeypatch):
+    root = tmp_path
+    (root / "A" / "youtube").mkdir(parents=True)
+    (root / "A" / "youtube" / f"T [{YT_ID}].opus").write_text("x")
+    monkeypatch.setattr(mf, "repair_file", lambda p, s, d: ["x"])
+    assert mf.repair_library(str(root), dry_run=False) == (1, 1)