v2.8.4

Fixed

• A Markdown note that is mostly one very large table no longer gets stopped at the per-file budget. When such a note had a code span anywhere and an unmatched ] in its rows, code detection rescanned the whole table at every ]. A note of 23 columns by 40,000 rows hit the 60 s budget as TIMEOUT during code-region detection; it now indexes in about 3 s. An unclosed [[ or <!-- in every row slowed link extraction the same way, and a code span beside a wiki link in every row did too. All of these now take time proportional to the note's size. Output for ordinary notes is unchanged. A note with a table of more than 5,000 cells is still parsed without the GFM table rules, and now also without inline links and inline HTML, so in such a note a backtick inside a link destination or an HTML tag can start a code span.
• gno doctor no longer reports a collection with no documents as a failed retrieval check (✗), and no longer exits 2 for it. The check is now informational (i) and names the collection with what to do: add files, then run gno update, which is now also the remediation command in the JSON output. A collection that has documents but fails the lexical proof is still an error. gno status, REST /api/status and the Web UI show empty collections the same way: activation.healthy stays true for them, gno status lists them as no documents yet (and reports Lexical activation: NO DOCUMENTS YET when no collection has documents), and they no longer produce pending connector proofs. MCP gno_status never reported them as a failure. Doctor checks can now have the status info.