matcher: token pre-filter for fuzzy name matching (6x faster startup)

The fuzzy fallback ran difflib against every catalog release (~11k) for each
unmatched disc, dominating library load (2.7s) and startup scans. Pre-filter
candidates to those sharing a significant (>=3 char) word with the disc
filename or PARAM.SFO title before the expensive comparison, and cache the
pre-computed word sets. Library load: 2.7s -> 0.44s; scan: ~0.5s.

app/matcher.py

@@ -38,7 +38,11 @@ def clear_cache():
 
 
 def _all_release_candidates(conn):
-    """Catalog releases + supplement entries with pre-normalized names."""
+    """Catalog releases + supplement entries with pre-normalized names.
+
+    Each entry is ``(normalized_name, significant_word_set, candidate)`` so the
+    fuzzy step can cheaply skip candidates that share no significant word.
+    """
     key = id(conn)
     cands = _releases_cache.get(key)
     if cands is not None:
@@ -53,7 +57,11 @@ def _all_release_candidates(conn):
         )
     ]
     cands.extend(_supplement_release(e) for e in _load_supplement())
-    cands = [(_normalize(c["name"]), c) for c in cands]
+    cands = [
+        (norm, frozenset(w for w in norm.split() if len(w) >= 3), c)
+        for c in cands
+        if (norm := _normalize(c["name"] or ""))
+    ]
     _releases_cache[key] = cands
     return cands
 
@@ -169,12 +177,21 @@ def match(conn, record):
     disk_norm = _normalize(record["name"])
     title_norm = _normalize(record.get("title") or "")
 
+    # Cheap pre-filter: only run the expensive difflib comparison against
+    # candidates sharing at least one significant (>=3 char) word with the
+    # disc's filename or PARAM.SFO title.
+    target_words = frozenset(w for w in disk_norm.split() if len(w) >= 3) | \
+        frozenset(w for w in title_norm.split() if len(w) >= 3)
+    if not target_words:
+        return "unknown", None
+
     scored = []
-    for cand_norm, cand in _all_release_candidates(conn):
+    for cand_norm, cand_words, cand in _all_release_candidates(conn):
-        ratio = max(
+        if not (target_words & cand_words):
-            SequenceMatcher(None, disk_norm, cand_norm).ratio(),
+            continue
-            SequenceMatcher(None, title_norm, cand_norm).ratio(),
+        ratio = SequenceMatcher(None, disk_norm, cand_norm).ratio()
-        )
+        if title_norm:
+            ratio = max(ratio, SequenceMatcher(None, title_norm, cand_norm).ratio())
         scored.append((ratio, cand))
     best_ratio = max((r for r, _ in scored), default=0.0)
     if best_ratio < 0.9: