feat: hard-exclude important user messages from compression

[ranxianglei]

Summary

• Add regex-based importance detector (lib/messages/importance-detector.ts) that classifies user messages as important using 5 categories of markers in both Chinese and English
• Hard-exclude important user messages from compression ranges — they survive intact in visible context, never enter summary blocks, never subject to GC truncation
• New config option: compress.protectImportantUserMessages (default: true)

Problem (issue #16)

Important user messages (instructions, corrections, constraints) could be compressed into summaries, then progressively lost across successive compression cycles. The "似是而非" (seems right but isn't) problem: content degrades with each compression round.

Solution

Two-stage classifier:

1. Gate: isLikelyDataInput() excludes logs, code blocks, diffs, tool output, structured data
2. Detect: 10 compiled regexes (5 zh + 5 en) check remaining natural language for:
   • Importance: 记住, 重要, 关键, remember this, crucial, essential...
   • Imperative: 必须, 务必, you must, ensure that, make sure to...
   • Correction: 不对, 错了, that's wrong, you misunderstood...
   • Reminder: 我之前说过, as I mentioned, like I said before...
   • Non-negotiable: 硬性要求, non-negotiable, deal breaker...

Code blocks are stripped before detection so keywords inside code don't false-positive.

Architecture

Follows the exact same hard-exclusion pattern as protected tools (PR #75):

• filterImportantUserMessages() in protected-content.ts — same structure as filterProtectedToolMessages()
• Applied in range.ts after filterProtectedToolMessages()
• Applied in message-utils.ts alongside protected tool check
• Config-driven: compress.protectImportantUserMessages: true (default on)

Test plan

• ✅ npm run typecheck — 0 errors
• ✅ 24 new tests in tests/importance-detector.test.ts
• ✅ All 610 existing tests pass (2 pre-existing Bun runner failures unrelated to this change)
• ✅ npm run build succeeds

Closes #16