A better similarity model, and a first run that cannot brick itself.
The checkpoint the map was built with has changed. Measured against the same
library, the old default agreed with itself far too much - every sound came out
nearly parallel to every other one, which is a map with no shape in it. The one
that replaces it is better on every measure that decides anything, and is
smaller. Libraries already built keep their analysis and are re-embedded once.
Fixed, and any of these could cost you a whole run:
- An interrupted download of the model left the app convinced it already had
it, unable to load it, and unable to fetch it again. It now knows the
difference between a model and part of one. - The analysis was thrown away if anything after it failed - a lost
connection, a model that would not load, or pressing stop. It is saved as
soon as it is done, so a second attempt only does what is left. - Stopping the app while it saved could corrupt what it had already analysed,
and the next run would quietly start again from nothing. - A failed download during setup reported itself as "was stopped", as though
you had done it - and sent you away from the log that said otherwise.
The model is now fetched while the environment is being set up, rather than in
the middle of your first build, and while a build is running it comes down
beside the analysis instead of after it. The window says which models are
already on disk, so a choice that costs a download says so first.
Every download size the app quoted was wrong. They are measured now: about 1GB
in total on Windows, 1.3GB on Linux.