It is expected that CBC results and MBC results do not lead to the same underlying estimated utilities and predicted choice probabilities. This has been examined and found multiple times and it is good that you have confirmed this once again. CBC asks the respondent to do something quite different from MBC: pick one pre-configured alternative from a set of pre-configured alternatives. MBC asks people to configure each item on the menu until an overall bundle has been constructed that the respondent prefers.
Rich Johnson, Jon Pinnell, and I found quite different results when comparing CBC to MBC in our paper we presented in 2006: "SIMULATING MARKET PREFERENCE WITH “BUILD YOUR OWN” DATA" downloadable from https://www.sawtoothsoftware.com/download/techpap/2006Proceedings.pdf
Also, Bakken also found in two earlier papers that the information provided or utilities estimated from comparable CBC and MBC attribute setups led to different utilities:
Bakken, David and Len Bayer (2001), “Increasing the Value of Choice-Based Conjoint with „Build Your Own‟ Configuration Questions,” Sawtooth Software Conference Proceedings, pp 99-110. downloadable from https://www.sawtoothsoftware.com/download/techpap/2001Proceedings.pdf
Rice, Jennifer and David Bakken (2006), “Estimating Attribute Level Utilities from „Design Your Own Product‟ Data—Chapter 3,” Sawtooth Software Conference Proceedings, pp 229-238.
downloadable from https://www.sawtoothsoftware.com/download/techpap/2006Proceedings.pdf
So, this is a strong argument that if the consumer is facing a decision where that person uses a menu to construct their preferred alternative, then the CBC approach will probably not mimic this appropriately and probably not lead to quite as good as predictions of actual market behavior as MBC. But, if the respondent selects one from many pre-configured product alternatives (and doesn't have the opportunity to multi-select and build her preferred product), then CBC would be the more appropriate conjoint tool to model preferences.