The problem
Three things about the sections bar, all of them visible the moment you look at
it.
There is no way to clear the markers. Sections can be added and deleted one
at a time. Starting over means clicking the small × on every marker in turn, and
an automatically generated set can easily be thirteen of them.
The same word appears twice. The bar carries an "Experimental - drag to
adjust." badge, and the control that produces the sections is also labelled
experimental. The badge also sits on the results rather than on the choice,
which is the wrong place for it: by the time you see the badge you have already
paid for the analysis.
Repeated kinds are indistinguishable. A song with two choruses shows
Chorus and Chorus, side by side. That reads as a duplicate or a bug, but it
is neither: boundaries and labels come from separate model heads, so two
neighbouring spans with one name are a real predicted boundary between chorus
one and chorus two.
The trap in the obvious fix
Merging adjacent sections that share a kind looks like the answer and is not. It
was the original behaviour, and it silently discarded five true boundaries on
the reference track before the UI ever saw them. Whatever fixes the display must
keep the boundary.
Constraints
- Clearing every marker cannot be undone, and an automatic set costs a whole
re-import to regenerate. There is no modal-confirm idiom anywhere in the app
to lean on.
- Every user-facing string goes through the i18n layer, in all nine tables, with
European Portuguese overriding only what genuinely differs.
- Removing the badge means removing what drove it too, or leaving dead state
behind that the next reader has to work out is unused.
The problem
Three things about the sections bar, all of them visible the moment you look at
it.
There is no way to clear the markers. Sections can be added and deleted one
at a time. Starting over means clicking the small × on every marker in turn, and
an automatically generated set can easily be thirteen of them.
The same word appears twice. The bar carries an "Experimental - drag to
adjust." badge, and the control that produces the sections is also labelled
experimental. The badge also sits on the results rather than on the choice,
which is the wrong place for it: by the time you see the badge you have already
paid for the analysis.
Repeated kinds are indistinguishable. A song with two choruses shows
ChorusandChorus, side by side. That reads as a duplicate or a bug, but itis neither: boundaries and labels come from separate model heads, so two
neighbouring spans with one name are a real predicted boundary between chorus
one and chorus two.
The trap in the obvious fix
Merging adjacent sections that share a kind looks like the answer and is not. It
was the original behaviour, and it silently discarded five true boundaries on
the reference track before the UI ever saw them. Whatever fixes the display must
keep the boundary.
Constraints
re-import to regenerate. There is no modal-confirm idiom anywhere in the app
to lean on.
European Portuguese overriding only what genuinely differs.
behind that the next reader has to work out is unused.