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.
Ilia Alshanetsky