Intermittent silent `ghc-pkg.exe` failures on `windows-latest` with GHC 9.2.8

There is, sadly, nothing much to be said about it: the log shows `ghc-pkg.exe` being run with various subcommands, then the test runner reports `Unexpected failure for `. Once it starts, it happens for all CI jobs for the next hour or so, then things return to normal.

It's probably not related to the recent switch of `windows-latest` to `windows-2025`, since there are both failing and succeeding runs on the new image (which went fully live on 2026-03-01). It has only happened with GHC 9.2.8, which does suggest there's some instability in that version that is somehow being triggered now but not earlier. It's also not a cache issue like the `GLIBC`/`GLIBCXX` symbol issue recently on Linux, since there have been failures since #11469 went in (see for example https://github.com/haskell/cabal/actions/runs/22730459904/job/65917786824#step:19:3870).

Start by comparing the failing and succeeding windows-latest CI jobs, focusing on the ghc-pkg.exe invocations and the test runner's Unexpected failure reports under GHC 9.2.8. Done means identifying the intermittent trigger and establishing a reliable CI fix or a clearly bounded workaround.

Written by the indexing model from the issue text.