fix(content): keep graphemes whole and excerpts fast

[m1ngsama]

Excerpts and summaries split emoji, and 40663fe made search excerpts about 10× slower.

• Excerpts and truncated summaries end on a grapheme boundary, so 👩‍💻 is no longer cut to 👩….
• Excerpt cost is back to the pre-40663fe level: 50 KB 5.2 → 0.66 ms, 300 KB 34.7 → 3.6 ms. Text is normalized in chunks that split only before printable ASCII or CJK U+4E00–9FFF, which NFKC never joins to what precedes them.
• Excerpts are identical to the current algorithm across 180,000 fuzzed queries mixing İ, fi, compatibility Jamo, ς/Σ, ZWJ and flag emoji, and across 990 excerpts from nbtca/documents.
• The list-marker pattern lives in one place; prose comments are cut to one-line traps.