I think they underestimated Cruz’s ability to turn out new voters. They used the old model turnout on him exclusively.
But, my gut also tells me we were locked into the old landline dominant polls that have proven wrong in England, Israel, and Kentucky. Only oldsters have exclusively landlines, too many don’t even have landlines, too many won’t answer their cellphones, so new methods are data collecting huge samples online, and THEN using random selections from that data to supplement the landline/cellphone data. It appears to work better.
And perhaps overestimated Trump's.
There could also be a bit of "confirmation bias" in the poll numbers -- pollsters hearing what they expected to hear, and reporting it.