I am pretty sure it is dmd regression - there have been several major changes in symbol generation recently. I have been trying to reduce the test case yesterday but got damn busy since then and won't be able to get back to it in a week at least.

2 most suspicious ones:
https://github.com/D-Programming-Language/dmd/pull/2550 has changed the placement of emitted symbols
https://github.com/D-Programming-Language/dmd/pull/2566 has changed the algorithm for anonymous lambda literal symbol generation