so rather than hacks of composite PEX_PATH w/ shadow and dupe problems galore, we’d love to offload the problem of dependency storage/retrieval to the build system/runtime for our ML cases and be able to build large pex environments in 1 go: hence the above thought.