v0.9.2: a value that wraps is still one value

• A Markdown frontmatter value that runs over several lines is now scanned whole. A key: > or key: | block scalar, a wrapped plain value and a - item list all carry the value on the lines after the key, and both engines read it from the key's own line only, so generated by ChatGPT on a continuation line survived a clean. #402 joins the continuation lines before the value check, in inspect_markdown and clean_markdown alike, and the clean drops the whole key as it already did for a hit on the first line. Blank lines and comments inside the block are skipped, and an agent definition's multi-line description is still checked.
• A > inside a quoted <meta> value no longer ends the tag. #401 makes the tag match skip quoted attribute values, so content="Compare Claude > GPT-4o" is read as one tag: the owner's own description is no longer flagged, nothing is left behind as stray text after a clean, and a marker behind the > is no longer hidden.
• _meta_attrs is a single forward scan. #407 replaces the backreference regex, which CodeQL flags as polynomial ReDoS on a tag from an uploaded document. The rules are the old ones: substring match on the attribute name (filename="x" still yields name), whitespace around =, an unquoted value or an unclosed quote is not read, and a later attribute wins.
• data-aid and data-airline are no longer taken for AI attributes. #400 lets the data-ai match continue only after a - or _, so the bare data-ai, data-ai-* and data-ai_* are still found and stripped and an ordinary application attribute is left alone.
• The web-extension strings added to classify_finding_confidence are ported verbatim. This port never produces those findings: the OOXML web-extension scan (#380, #394) is not ported.
• 1865 tests, twelve upstream anchors, three of them moved to upstream 5c18093.

Ports upstream guillaumemeyer/watermarks-remover at 5c18093. Closes #33.