Right now we create chunks and then apply platform selection in two separate stages, which is sub-optimal.

For example, consider this situation.
Groups 1, 2 and 3. Platform A, B and C.
Groups 1 and 2 are platform independent.
Group 3 is platform-dependent and must run on all platforms.

If we create chunks and apply platform selection later, we end up with three chunks: a chunk with 1, 2 and 3 for each platform. In the end, we run all groups three times.
If we create chunks and apply platform selection at the same time, we end up with three chunks: a chunk with 1, 2 and 3 on platform A, a chunk with 3 on platform B, a chunk with 3 on platform C. In the end, we run two groups once and one group three times.