This is a proposed fix for #3741.
(I stumbled upon this when using pragma legacy_alter_table=1 as a workaround for #1686, after I had added a few rows to a table.)

When saving, the GUI thread blocks on DBBrowserDB::waitForDbRelease line 847, where it waits for db_used to become false. This must be done by ~db_pointer_type (DatabaseReleaser).

In RunSql::executeNextStatement this deleter is normally invoked by RunSql::releaseDbAccess, but in the case of a pragma statement on a dirty DB, the function exits early (line 140) without calling releaseDbAccess. This creates a deadlock, as db_used is never set to false - thus hindering the GUI thread from accepting any user input.

Calling the aforementioned method before returning resolved the issue for me. I'm neither familiar with the sqlitebrowser codebase nor with C++ in general, so this probably isn't the best solution.