Hmmmmmmmmmmmmm...... That was a slog! :-)

My take:

I buy the "sensible defaults" argument most: bracket should "do the generally right thing"
Sure... fix hClose (I can't tell from the discussion if there is or isn't really a bug there...but please, don't tell me here - that isn't the discussion at hand) - but I don't see how even in the presence of fixed library functions, bracket still shouldn't "do the right thing".
As always, I side with Gregory on the deep caution stance - I like things to be stable for the long haul....
...but in this case, I didn't see anyone point to any preexisting code that would break under this proposal. Perhaps we havent done enough auditing.... but no one in this long thread pointed at even one such case. (Gregory - yours feels a bit contrived, even if you've seen it is some other, non-Haskell, proprietary system :-) ) Whereas John Lato has given us a simple, reasonable, but utterly broken example under the current bracket.

So here's how I vote:

+1 for Merijn's Proposal - but I'm all for as much aduiting and testing before we go as we can stand

-0.5 for interruptibleBracket - because it feels like "we weren't sure if anyone needed this, so we kept it for completeness..." and then it just becomes something to maintain.