fix: add Chinese noise filter patterns

[PanYuxn]

Closes #48

Problem

noise-filter.ts only contains English patterns. Chinese greetings (你好), denials (我不记得), meta-questions (你还记得吗), and acknowledgments (好的, 谢谢) pass through unfiltered and get stored as memories — polluting retrieval results over time.

This is especially impactful with autoCapture enabled, as every 好的 and 谢谢 becomes a permanent memory entry.

Solution

Add Chinese patterns to all three filter categories:

Category	Examples	Count
Denial	我不记得, 找不到相关记忆, 我没有相关信息	8 patterns
Meta-question	你还记得吗, 我之前提到过, 我有没有说过	6 patterns
Boilerplate	你好, 早上好, 好的, 谢谢, 知道了	7+2 patterns

False-positive prevention

Chinese acknowledgment words (好的, 谢谢) often appear at the start of meaningful sentences:

• 好的方案是使用Redis做缓存层 — should NOT be filtered
• 谢谢分享，我觉得这个思路很好 — should NOT be filtered

Solution: length-gated SHORT_BOILERPLATE_PATTERNS — prefix matches like 好的/谢谢 are only treated as noise when total text length ≤ 10 characters. Short = filler, long = real content.

Test Results

52/52 tests pass:

• 38 noise cases correctly filtered (9 denial + 7 meta-question + 19 boilerplate + 3 English)
• 8 false-positive guards all pass (meaningful text kept)
• 6 integration test assertions pass

Changes

• src/noise-filter.ts — Add Chinese patterns + SHORT_BOILERPLATE_PATTERNS with length gate
• test/noise-filter-chinese.mjs — New test: 52 cases covering all categories + false-positive protection

No new dependencies. Existing npm test passes.

[AliceLJY]

PR #49 Review Summary

Tested on Gateway 2026.3.2 + claude-opus-4-6. Unit tests 52/52 pass, runtime memory_store verified on Discord.

One issue to fix before merge:

/你知道.*吗/ in META_QUESTION_PATTERNS is too broad — it catches legitimate questions like "你知道Redis的持久化策略吗". Suggest narrowing to /你知道我(说过|提过|告诉).*吗/ or removing it (the other 5 Chinese meta-question patterns already cover the intent).

Everything else looks good. The SHORT_BOILERPLATE + length-gate design is clever and well-tested. Approve after the regex fix.