Keywords starting/ending with non-word characters (brackets, parens,
dashes) never matched because \b requires a word↔non-word transition.
Now uses smart boundaries: \b for word-char edges, (?<=\s|^)/(?=\s|$)
for non-word edges. Also pre-compiles regex patterns once per batch
instead of recreating them for every description × keyword combination.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>