feat: add preservePatterns for domain-specific preservation

[SimplyLiz]

Summary

• Adds preservePatterns option to CompressOptions — users supply regex patterns that force hard T0 preservation, enabling domain-specific content protection (legal, medical, academic) without the heuristic classifier
• Patterns are checked after built-in classification but before JSON detection, consistent with the existing rule priority chain
• Tracks messages_pattern_preserved stat in CompressResult.compression

Changes

File	Change
src/types.ts	New option field + new stat field
src/compress.ts	Thread patterns into classifyAll, track stat
tests/compress.test.ts	7 new tests (match, no-match, multi-pattern, empty array, code-split precedence, dedup precedence, tokenBudget integration)
docs/api-reference.md	Document option and stat
docs/preservation-rules.md	Priority table update, new section with legal/medical/academic examples

Test plan

• All 343 tests pass (npm test)
• Lint clean (npm run lint)
• Format clean (npm run format:check)
• Benchmarks pass with no regression (npm run bench)