This is a common scenario, but interestingly enough, it is never addressed in texts/manuals on statistics. This may have a significant impact on the concept of rational subgroups and the warning against mixing process streams, but no guidance is given on the appropriate way to handle it.
So here are my thoughts on the subject. To keep it simple, we will continue your 2-cavity mold scenario. The two cavities will not be identical, and therefore may represent two process streams. If they are indeed different, they should not be included in the same subgroup and should not be combined into the same capability study.
My recommendation would be to collect samples from each cavity and perform a 2-sample t-test on the cavity means. If the t-test is statistically significant, treat the cavities as different process streams. If the test is NOT statistically significant, treat them as a single process stream. You may safely include them in the same subgroups.
The next thing that I would consider for your specific question is, what is the major source of variation? Is it the variation between the two cavities? Or is it the variation from heat to heat? The samples should focus on the major source of variation that has the potential to vary over time. The difference between cavities will likely not vary until a long time passes, while heat to heat variation may occur over a much shorter time frame.