Fix bsearch to never return the right endpoint (in case we can't-split)

[pavpanchekha]

Fix #840. This was a very subtle bug. Basically, suppose regimes narrows our split point to somewhere in [a, b], but also suppose we can't split after b. Then bsearch picks something in [a, b]. But it can't be b! So now we check for b and in that case return a.