Yes, glue logic that maintained a n-level write-back queue would be a beauty of "simplicity" to implement - especially if it also needs to have a look-ahead snoop in case a read is made of an entry still in the pipeline :)

In a very large percentage of the cases where a microcontroller needs access to a lot of RAM, the stored information has some form of structure where it can be seen as records, and where the program need not play with direct accesses but can instead implement get_record() and put_record() functions. This would directly remove the need for MOVX instructions, and allow a serial interface to be used without any hw complexity and with a very clean sw implementation. And SPI-connected FRAM is rather fast too.