Here are my thoughts, but I warn you my thinking of the use of blocks may be different than the typical enumerative approach. The purpose of the RCBD is two fold:
1. To partition the effects of noise so as to increase inference space while possibly improving the precision of the design, and
2. To assign the effects of the block AND block by factor interactions. The later being of utmost importance for robust design.
Regarding your query, I think it is unreasonable to assume the effects of factors will remain constant over changing noise (this quantified by block-by-factor interactions). There is so much empirical evidence this is not true, I'm not sure why anyone would assume it is true. Why do products not perform as intended in the hands of the customer? (Why do companies spend so much money on reacting to customer complaints and warranty?). Why due process settings for an injection molding machine not remain constant the next time the machine is setup? Why do manufacturing companies have to change their process when new batches of raw materials are introduced to the process?
Back to your question, this is an indication your factor(s) is not robust to noise. You will need to iterate to identify which specific noise x's might be the culprits. Own one hand, this is great news. You identified the results of your experiment run under the conditions of the first block do not repeat when you run under the conditions of the second block. Celebrate you identified this and you can continue your investigation to uncover the causal structure. Think of it this way, what if you did not block and went with the results of the first block conditions. Your model would not work when conditions change, which invariably they will.
"All models are wrong, some are useful" G.E.P. Box
