Walter,  Jody and I were discussing this topic just a few days ago.
Your letter is very timely.  Here are some thoughts of mine...

The conversion tasks may be grouped like this:

1. transcribing the music roll to a "roll-image" disk file.¶
2. converting the roll-image file into a MIDI file¶
3. converting expression coding into MIDI velocities.

The steps above are the processes I use, which I have adapted, with little
modification, from the processes developed by Richard Tonnesen and Wayne
Stahnke.  Briefly stated, the ultimate goal of the transcription process (#1) is
to produce a disk file representing the image of the music roll,  with
sufficient precision that an exact duplicate of the paper roll made be created
(perforated).

It is very convenient (although not required) to convert the file format to
MIDI, in order to perform the editing using already-published editor
programs.  The conversion to-and-from MIDI should not degrade the
precision of the roll-image file.  As you mention, it is important to properly
extend note durations for the MIDI file, and then truncate the MIDI durations
properly when converting back to image format (assuming that you want to
perforate a roll from the edited MIDI file).

Once the music is in MIDI file format the expression coding channels may be
"interpreted" to yield MIDI velocities, which are appropriately applied to
each note event.  I wrote an off-line program for this purpose, so that if I
change the expression coding (editing the Ampico crescendo, for example) I
can simply re-interpret the new coding and hear the result quickly.

I hope to prepare a sample, for our group, of an Ampico roll in MIDI format
with interpreted expression, accompanied by a description of all the
processes.   In the meantime, I hope there will be ideas forthcoming from
our members.  This is a very interesting topic.

-- Robbie Rhodes