Fixed the bugs introduced by the last commit! Output is now *identical* to that
produced by the old version, based on a-lists.
text {*
The Isar toplevel works differently for interactive developments
vs.\ batch processing of theory sources. For example, diagnostic
commands produce a warning batch mode, because they are considered
alien to the final theory document being produced eventually.
Moreover, full @{text undo} with intermediate checkpoints to protect
against destroying theories accidentally are limited to interactive
mode. In batch mode there is only a single strictly linear stream
of potentially desctructive theory transformations.
*}
\item @{ML Toplevel.empty} is an empty transition; the Isar command
dispatcher internally applies @{ML Toplevel.name} (for the command)
name and @{ML Toplevel.position} for the source position.