v0.18.0

Every document converted from Google's HTML export, which is every document
with an image, keeps its bold, italic and struck text, its tables' marks and
links, and loses Google's wrappers in a merged table. The MCP server stores and
returns a document without its front matter, fetch names its folders and its
source, and the first measurement of search sets its defaults; see
ADR-042,
still proposed.

Changed

• Bold, italic and struck text from Google's HTML export. Google draws
  them with classes of its stylesheet, which the converter removed without
  reading, so every document with an image lost every mark. The rules that
  name one class are read first, and a span that draws a mark becomes
  **, * or ~~. A heading's bold is not repeated, marks Google splits
  over several spans are joined, and where Markdown would not read the
  delimiters (punctuation inside the mark and a letter or another mark beyond
  it) the mark is written as its HTML element.
• Markdown tables keep their cells' marks, links and paragraphs. A cell
  was its plain text: marks and links were lost, and paragraphs ran together.
  A cell's paragraphs and line breaks are now <br>, and a pipe is escaped.
• Tables with merged cells keep their cells without Google's paragraph
  and span wrappers, its no-break spaces and its default colspan="1" rowspan="1".
• The converter version is hybrid-v4, so the next sync exports every
  document again, once.
• The MCP server keeps each document without its front matter. The copy
  in R2, and so the index, no longer holds the sync's bookkeeping; the
  catalog hashes that text, so a sync that changes only the front matter
  rewrites nothing. The content health page measures the same text.
• fetch adds path, the folders as a list, and source, the Google Doc
  or PDF in Drive, to its metadata. A document cut at mcp.fetchCharacters
  ends with a line saying it continues and where the whole of it is, and is
  never cut inside a character.
• search cuts a passage with no sentence end between words, not inside
  a word or a tag.
• mcp.search.contextChunks defaults to 0. With AI Search's default
  1,024-token chunks, neighbouring chunks changed no ranking and only added
  text the Worker cut away.
• mcp.search.vectorThreshold is described as it behaves: AI Search ignores
  it while reranking is on.
• ADR-042 records the first measurement of search: 34 questions in English
  and Russian, one setting varied at a time, and 512-token chunks on a second
  instance. ADR-043 measures the text fetch returns. The runbook says how
  to see documents an index left in error or running.

Upgrade note

Bump the packages and the workflow pins. The next sync is a full export
(hybrid-v4), and a targeted sync is refused until it has run. After the
deploy that follows, the MCP server rewrites every document in its bucket
once and AI Search reindexes them. A project that set
mcp.search.contextChunks to 0 may drop the line.