I'm sorry to keep banging on the definition distribution bell, but here is another (hopefully well-focused enough) question on that vein.
If one is trying to distribute the definitions in a package to the subkernels that will perform a parallel computation, there are several options for it, most prominently
ParallelNeeds. One of the features explained in their documentation is that the definitions they distribute are made available to any new kernels that might get started after the distribution. This could be because a kernel dies and is respawned halfway through a calculation, or simply because
DistributeDefinitions gets called before
In particular, then, the documentation promises that
DistributeDefinitionsin effect "registers" the definitions for the symbols $s_i$ so that they are automatically distributed to every new parallel kernel that may be launched.
ParallelNeedsin effect "registers" Needs specifications so that they are automatically performed in every new parallel kernel that may be launched.
This is in contrast to
ParallelEvaluate[Needs[...]], which as opposed to
ParallelNeeds will (i) not start new kernels, and (ii) not pass the definitions to any kernels started after the command is evaluated.
The documentation, though, is pretty mysterious on the details of how these two functions manage to "register" the stuff they're distributing; my question, therefore, is how this happens internally. To put this on a more concrete, footing, though, I'm happy to accept as an answer a description of the 'register' where the definitions are stored, and how it could be accessed by the user.