v0.19.0

Added

• Database context for MySQL and ClickHouse connections, which browse a whole
  server rather than one database. The editor breadcrumb reads
  connection / database / table, where the database is a picker showing what an
  unqualified FROM users resolves against; browsing a table from the tree moves
  the tab into that table's database. The target is per tab, so a split view can
  compare two databases over one connection.
• Right-click menu in the schema tree: browse a table, open a query bound to a
  namespace, set the active database, and copy a plain or qualified name.
• Connections can be marked Local, Dev, Staging, or Prod, and the write guards
  follow the marker. Local asks about nothing short of destroying an object, so a
  scratch database stays out of your way; Prod confirms from the first risky
  statement, makes every confirmation typed rather than clicked, and offers no
  "Don't ask again" at all. Unmarked connections behave exactly as before. The
  marker shows as a coloured badge on the connection card and beside the
  connection name, and the connection form guesses one from the host and name.
• Optional second opinion from your AI agent on a statement a confirmation has
  stopped, for the mistakes a keyword check cannot see: an inverted filter, a
  value that reads wrong against the schema, a join that fans out. Off by
  default, since enabling it sends the statement and a schema summary to the
  configured provider. It is advice and nothing more: it runs nothing, reads no
  rows, and can never unlock or shorten a confirmation, so a slow or mistaken
  model costs you a line of text rather than a table. Settings → Query → Safety.
• The confirmation for a destructive statement now asks the database how much it
  will touch: "This affects 8,412 rows", or "orders holds 8,412 rows" before a
  DROP TABLE. The count runs over the same table and WHERE clause the
  statement uses, so a predicate that matches more than you expected shows up
  before the run rather than after. It never holds up the dialog: it fills in a
  moment later, gives up after two seconds, and says plainly when it can't tell.

Changed

• Statements are graded before they run, and the confirmation matches how
  dangerous they actually are. A filtered UPDATE ... WHERE id = 42 or an
  ALTER TABLE ... ADD COLUMN runs uninterrupted; an UPDATE or DELETE with
  no WHERE, a privilege change, or a MERGE stops and says what it noticed
  ("No WHERE clause: this removes every row in orders"); a DROP TABLE or
  TRUNCATE asks you to type the object's name before the run button enables.
  GRANT/REVOKE, CREATE USER, stored-procedure calls,
  ALTER TABLE ... DROP COLUMN, and a WITH query whose CTE writes are now
  caught too, where they used to run unasked.
• "Don't ask again" now silences only the kind of statement in front of you, so
  hiding the routine prompts no longer hides the one before a DROP DATABASE.
  The query.confirm_destructive setting is replaced by query.confirm_from,
  which takes write, risky, critical, or never and is migrated on first
  launch; confirm_destructive = false becomes critical rather than never,
  since it used to be the only way to stop being asked about ordinary writes.

Fixed

• A multi-statement script run from the editor (a CREATE TABLE followed by its
  indexes and seed rows, say) executed only its first statement and reported
  success on SQLite, and failed with a bare engine error elsewhere. The whole
  script now runs as one transaction, and the toast reports how many statements
  committed.
• An unqualified query on a MySQL connection that named no database reported the
  server's bare "No database selected" with no way to act on it. The results pane
  now explains what happened and lists the databases on the server, so picking
  one sets the target and re-runs the query.
• ClickHouse connections that named no database silently resolved unqualified
  queries against default, even while browsing another database, which could
  return a same-named table's rows rather than an error. The database is now
  explicit and selectable.