After being initially skeptical I was convinced by a colleague that this is a good model.

The merge commit acts as a kind of "pushlog"; it tells you what actually landed together as a single unit. It probably also tells you what passed CI; although many projects state that you shouldn't have individual commits that don't pass CI it's rare that this is enforced below the level of the PR. That should be good for bisection. Because the commits are always rebased onto master (and of course you never allow merge commits other than from code being integrated into master) your history is relatively clean, you don't get multiple overlapping branches, but something like

  M1 ------- M2 -------- M3
   \ B1 - B2 / \ C1 - C2 /

Because the merges are --ff-only you don't get the confusing situation where the merge commits themselves contain changes.

In principle this seems like the ideal way to use the tools git provides. However I understand there are some minor rough edges e.g. with magic needed to tell git bisect how to only try the merge commits.