fix(normalize): strip noise blocks with multi-paragraph bodies and indented tags

[mazurd-acre]

Problem

Two escape paths let Claude Code harness chrome leak into mined drawers and pollute search/recall:

1. <task-notification> and <system-reminder> blocks carrying multi-paragraph payloads (subagent results, recalled memories) were never stripped: the tag-body pattern refused to cross blank lines, so exactly the largest noise blocks survived wholesale.
2. Claude Code emits slash-command chrome with indented tags (<command-message>, <command-args> after <command-name>), which the line-start anchor missed. command-args was also missing from the noise list entirely.

Found by mining ~90 real Claude Code sessions: thousands of drawers contained raw task-notification payloads, and semantic recall surfaced harness boilerplate instead of content.

Fix

• Tag bodies may now span blank lines, guarded by a same-tag-reopen lookahead instead of the blank-line rule: a dangling open tag still cannot merge with a later block and eat the real content between them (the original span-eating protection this rule existed for).
• The line anchor tolerates leading whitespace after the optional > blockquote marker.
• command-args added to _NOISE_TAGS.

Testing

• 3 new regression tests (multi-paragraph block, dangling-tag safety, indented slash-command chrome).
• Full test suite green.
• Re-mined the same ~90 real sessions (20k+ drawers): zero noise blocks survive, inline prose mentioning the tags is still preserved.

[fatkobra]

Blocking: allowing the noise-block body to cross blank lines reintroduces a destructive dangling-tag case.

The negative lookahead prevents an unclosed opener from merging with a later opening of the same tag, but it does not prevent that opener from consuming arbitrary transcript content until a later standalone or literal closing tag.

For example, an unclosed <system-reminder> followed by legitimate conversation text and a later user/code example containing </system-reminder> can cause all intervening content to be removed.

That contradicts the conservative normalization invariant: when pairing is uncertain, original transcript content should survive.

Please bound matching using actual transcript/message boundaries, or replace the cross-paragraph regex with a small parser that strips only demonstrably paired noise blocks.

Please add a regression containing:

1. a dangling opening tag;
2. blank lines;
3. legitimate user/assistant content; and
4. a later closing-tag token without another opening tag.

The legitimate intervening content must remain unchanged.