Two things; 1) src_test's default of addpredict / , and wiping any further addpredict's the ebuild adds during src_test is undocumented. All PMs do this thus it's safe to document it, and should be documented since it's behaviour ebuilds rely on. 2) pkgcore doesn't actually save SANDBOX_* vars between phases. This was an intentional choice long, long ago, at the time to curb the var from building up into a rats nest. Rules changed since then, never caught wind of that, frankly haven't seen reports nor breakage up until recently. Scanning the tree, pretty much the sole "carry the settings forward" that I'm aware of is java-utils's usage for adding blocking of crazy shit like "don't go writing to system coredump settings you tool". Everywhere else, looking through it the writes/predictions are contained to just the phase that has it. Minimally, we should expand the spec to be explicit about this; in the same instant, I'm strongly of the view we should probably make sandbox manipulations contained to the originating phase unless an option is given explicitly forcing it to persist.