Lataa Data Maskit 0.12.0

更新摘要 / Summary
- 数据库密码、API 密钥这类配置项现在能正确识别带前缀的写法，以前只认裸关键词。
- 日志页现在能分清「模型没提到敏感值」和「真的没还原」，不再把两种情况都显示成同一个状态。
- 新增「只提醒」「拦截」两种处理方式和值级白名单，误报可以只放行那一个值，不必关掉整条规则。
- Credentials written the way .env files actually look are now detected; previously only the bare keyword was.
- The log page now tells "the model never mentioned the value" apart from "the value was not restored", instead of showing one status for both.
- Added "warn only" and "block" per-rule handling plus a per-value allowlist, so a false positive can be released without disabling the whole rule.
修复
- 凭据识别覆盖 .env 真实写法：DB_PASSWORD / MYSQL_ROOT_PASSWORD / OPENAI_API_KEY / db.password 等带前缀或点号的键名以前全部漏检，现在正常脱敏。
- 工具定义区不再被改写：工具名、命名空间、参数结构、语法约束以前会被打码，导致上游拒绝请求；现在只扫描其中的说明文字。
- 视觉等价写法不再绕过：零宽字符插入、全角数字与符号（１３８…、alice＠example.com）以前整条漏检，现在能识别，且发出去的内容与用户输入逐字节一致。
- 已知敏感值不再重复出网：某个值一旦被确认敏感，之后在任何位置出现都会被替换，不必再靠 password= 这类锚点。
- 新增出口前复检：脱敏后仍能匹配到原文的请求会被拦下，不再上行。
- 新增中文姓名补漏与剔误：客户陈建国 这类句式不再漏检，张开的括号 这类普通词不再被当成人名。
- 还原状态细分：no_placeholder_in_response 拆成「模型没提」「模型提了但形态坏了」「响应没走完」，日志页可区分。
- 「只提醒」「拦截」两种处理方式现在对所有规则都生效（此前只对部分规则有效），且放行记录会显示在日志里。
新增
- 每条规则可选「脱敏 / 只提醒 / 拦截」三种处理方式。
- 值级白名单：可放行单个具体值，大小写不敏感，最多 200 条。
- 脱敏取证：可在日志页查看最近 50 条「实际发出去的内容」，仅内存、不落盘。
- 高熵凭据检测（默认关闭）：覆盖没有厂商前缀、也没有赋值标签的内部服务令牌。
Fixed
- Credential detection now covers real .env key names (DB_PASSWORD, MYSQL_ROOT_PASSWORD, OPENAI_API_KEY, db.password), which were previously missed entirely.
- Tool definitions are no longer rewritten: tool names, namespaces, parameter schemas and grammar constraints used to be masked, causing upstream rejections; only their descriptions are scanned now.
- Visually equivalent spellings no longer bypass masking: zero-width insertions and full-width digits/symbols (１３８…, alice＠example.com) were missed entirely; they are now detected, and what leaves the machine stays byte-identical to what the user typed.
- A value once confirmed sensitive is now masked wherever it appears, without needing a password=-style anchor.
- New pre-egress re-check: a request that still matches an original value after masking is rejected instead of being sent upstream.
- Chinese name completion and rejection: "role word + name" phrases (customer, contact, author …) are no longer missed, and ordinary words that merely start with a surname are no longer treated as names.
- Restore status is now split: no_placeholder_in_response becomes "the model never mentioned it", "the model rewrote the placeholder", or "the response did not finish".
- "Warn only" and "block" now apply to every rule (previously only some), and what was let through is shown in the log.
Added
- Each rule can be set to mask, warn only, or block.
- Per-value allowlist: release a single concrete value (case-insensitive, up to 200 entries).
- Masking forensics: the log page can show the last 50 bodies actually sent upstream (memory only, never written to disk).
- High-entropy credential detection (off by default) for internal service tokens with neither a vendor prefix nor an assignment label.

扩展更新提示 / Extension update notice
✅ 本次发版扩展无改动：已安装的扩展无需重新加载。
✅ The browser extension did not change: no need to reload it.
Full Changelog: https://github.com/xiaYuTian11/maskit/compare/v0.11.4...v0.12.0