docs: publish v2.3.0 documentation and release notes (#2192) * Add versioned docs for v2.3 * docs: publish v2.3.0 release notes and announcement * docs: remove the temporary local build configuration * docs: limit the LLM bundle to the latest released version --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Phillip LeBlanc <879445+phillipleblanc@users.noreply.github.com>

docs: note query failure logs, Iceberg append layout, and Cayenne cross-batch keys (#2307) * docs: note query failure logs, Iceberg append layout, and Cayenne cross-batch keys Document operator checks that match the released runtime: query_failures codes and where the error text is returned, how Iceberg append polls skip files and row groups, and Cayenne's per-batch primary-key upsert. Co-authored-by: Luke Kim <lukekim@users.noreply.github.com> * docs: correct Cayenne no-on_conflict and err_code claims against v2.3.2 - Cayenne: the in-batch "keep the first row" and stored-key "incoming row is dropped" rules are on_conflict: drop, not a primary key with no on_conflict. On spiced v2.3.2's append refresh, no on_conflict kept the last in-batch copy and replaced a stored key, while drop kept the first copy and kept the stored row. Scope the check to an append refresh: a refresh_mode: full load stored both copies of a duplicated key. The cross-batch rejection applies with or without on_conflict. - Troubleshooting: a parse error is SyntaxError, but an unknown table or column (and a GROUP BY violation) is labeled InternalError, not QueryPlanningError. ErrorCode::from has no arm for DataFusion's Diagnostic wrapper, so those planning errors fall through. * docs: address Copilot review of #2307 and mirror the guidance into 2.3.x - constraints and datasets reference: carve out Spice Cayenne, whose append refresh collapses a key repeated within one batch (upsert keeps the last copy, drop the first) and rejects a key repeated across batches, instead of failing on an in-batch duplicate. - Iceberg: bin-pack compaction can mix old and new rows rather than always mixing them; a sort order applies within partitions, so the benchmark covers a time partition, a sort order, or both; on Cayenne's no-primary-key log, rank the latest row per key before the soft-delete filter (filtering first resurfaces the previous live version). - Mirror the additions into versioned_docs/version-2.3.x, which the site serves as the latest docs; the guidance describes v2.3.2 behavior. The append-only-log paragraph stays trunk-only because 2.3.x has no such section. --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Luke Kim <lukekim@users.noreply.github.com>