Bug#68585: looks like it applies holds too late

[Adam Borowski <kilobyte@angband.pl>]

Looking more closely (because it's especially bad for multiarch), I see that
it appears to be caused by applying holds too late.

Let's say we have the following versions and dependencies:
  A=1 (installed)
  A=2 Breaks: B
  B (installed, held)
  C (installed) Depends: B

(or any similar scenario, in my case A having available versions A:amd64-2
and A:i386-1, B:i386 depending on A:i386)


If the resolver wants to upgrade A to version 2, it will decide that it
needs to remove B and C.  It only then processes holds, marking B and
(transitively) A as kept.  C still remains marked for removal, even though
any reason to do so is gone.

-- 
ᛊᚨᚾᛁᛏᚣ᛫ᛁᛊ᛫ᚠᛟᚱ᛫ᚦᛖ᛫ᚹᛖᚨᚲ