fix+perf: lettered sub-items, ~10x faster unmask (v2.6.3, v2.6.4)

[click0]

Two deferred audit items closed

2.6.3 — false-positive initials

• п. В. Петренко (clause item B), ст. А., абз. Б., див. п. В. no longer masked as PIB
• When a service abbreviation (п., пп., ч., ст., абз., гл., розд., …) directly precedes an initials+surname pattern, the leading letter is a clause marker, not a name
• Real PIB after ordinary words still masked
• Backfilled CHANGELOG entries for 2.6.1 / 2.6.2

2.6.4 — unmask performance

• unmask_other_data scanned the full text once per mask — O(masks × text)
• Now one alternation regex (longer masks first), single text walk
• Per-mask occurrence counter keeps instance tracking exact for collisions (two originals → same mask)
• 278 KB doc: ~30 s → ~2.8 s (~10×), roundtrip verified
• Slow per-mask path kept as fallback for masks that break regex compilation

Test plan

• 433 tests pass (+6 new: lettered sub-items, collision, substring priority, case, empty)

https://claude.ai/code/session_01XT6iUWaQgahXDB9TWX9Bq7

---

Generated by Claude Code