Seems there are various things being mixed together here.
First is the robustness issue: if the control directory cannot be created, we had better report that properly, and fail.
JENKINS-28400 covers this.
Then there is an issue reported with DinD, which is perhaps best avoided by using bind-mounts of the Docker socket; PR 31 purports to address this.
Then there was a comment about containers with an entry point. If reproducible, clearly a distinct bug.
Then there is some discussion of UID mismatches, which PR 25 might address, though there is no information given there about how to reproduce.