Two changes that close the "everything lives in one global pool" gap.
UI: Saved Queries sidebar now defaults to "This connection" and lists
only entries with `connection_id == activeConnectionId` plus unattached
entries (legacy global saves before F2). A small toggle ("All") above
the search box brings the previous behavior back when copying queries
between databases. Each row in "All" mode shows a tag with the source
connection's name; legacy global entries show "unattached".
Backend: delete_connection now best-effort cleans persisted state for
that connection — removes `memory/<id>.md` (delete_memory_core in
memory.rs) and drops every entry in saved_queries.json with the
matching `connection_id` (delete_by_connection_core in
saved_queries.rs). Entries with `connection_id == None` are deliberately
preserved. Cleanup errors are logged but don't block the deletion since
connections.json is the source of truth.
Memory was already per-connection (F1); query history already filters
by connection. This commit makes saved queries behave the same and
stops orphan files from accumulating.
Aleksey Shakhmatov
95c9470411