A soft fork adds limitations that make previously-valid blocks invalid, so miners who aren't enforcing _any_ limitations are effectively on the "didn't upgrade" side. But a hard fork makes blocks that were previously invalid valid, so if 75% are advertising big-block support but some miners who haven't advertised that aren't validating at all, you'll actually have more than 75% mining the big-block main chain.
PS. In reality you'll have much more than 75% advertising big-block support once the switch-over comes (if it ever does) because most miners will upgrade during the 2-week grace period rather than be left mining dead coins.
PPS. After losing money this way I doubt many miners will continue running with verification completely switched off; They have much better alternatives either by getting blocks propagated faster and mining normally or by at least verifying in parallel and giving up trying to mine on top of blocks once they've found that they're invalid.
Fair enough, but don't forget MPEX, etc. plan to sabotage any attempt to raise max_blocksize. IE, they will falsely advertise >1MB compliance in order to (prematurely) bait the hard fork into a prearranged GavinCoin Short kill box.
This is not a serious suggestion. You'd need 25% of mining power to pull it off even if zero miners switched in the grace period, but since most of the remaining miners will switch during the grace period you need more like [conservatively] 45%. MPEX control approximately 0% of mining power, and if they were able to acquire more than 25% they could block bigger blocks the normal way by refusing to upgrade.