perf: optimize dedup, parallelize handle searches and enrichment

The dedup hot path recomputed normalize_text() 4 times per comparison
and recomputed item_text() on every inner-loop iteration. Pre-computing
n-gram sets and token sets into a _PreparedText cache cuts dedup time
by 6x (2.16s to 0.39s on 300 unique items).

Bird handle searches spawned one Node process per handle sequentially.
Now uses ThreadPoolExecutor so N handles run concurrently. Same pattern
applied to YouTube comment enrichment (was serial, Reddit was already
parallel) and the retry-thin-sources phase in the pipeline.

Clustering now pre-computes candidate text and uses prepared_similarity
for the O(n^2) grouping and MMR representative selection loops.

Minor: _is_wsl() cached with lru_cache, Bundle.add_items() uses
extend() instead of list concatenation.

End-to-end: 5.2s -> 3.7s (29% faster) on a typical 4-source query.

scripts/lib/cookie_extract.py

@@ -7,6 +7,7 @@ Only uses Python stdlib — no external dependencies.
 """
 
 import configparser
+import functools
 import logging
 import platform
 import shutil
@@ -18,8 +19,12 @@ from typing import Dict, List, Optional
 logger = logging.getLogger(__name__)
 
 
+@functools.lru_cache(maxsize=1)
 def _is_wsl() -> bool:
-    """Detect if running under Windows Subsystem for Linux."""
+    """Detect if running under Windows Subsystem for Linux.
+
+    Cached after the first call since /proc/version doesn't change at runtime.
+    """
     try:
         return "microsoft" in Path("/proc/version").read_text().lower()
     except OSError: