What is expected from Britney's arch:all handling?

 * arch:all packages must be installable on at least 1 architecture.
 * arch:all packages may be uninstallable on a strict subset of the architectures assuming they have no arch:any rdeps on that architecture.
 * arch:all that used to be installable must not regress without a hint/manual approval except:
   * If the source previously provided arch:any binaries but no longer does for that architecture.  (Assumption, if arch:any + arch:all binaries are provided, the arch:all binaries are assumed only to be relevant on architectures with arch:any binaries)
     * This has an exception itself:  If all arch:any packages are dropped, then regressions should not be allowed.

Currently, dak decides which arch:all package is available on which architecture. Every arch:all package that is put into testing by britney is handled the same way by dak. Maybe this decision shouldn't be handled by dak, but by britney. This would require changes to the interface that britney uses and probably also changes to the projectb database.
 - Assuming Packages-all becomes a thing, then we still cannot do it per architecture, so we should probably not assume we can control arch:all per architecture.

Notable cases:
 * Drop of every arch:any leaving only arch:all.  This should ideally not regress