Reviewing for CCB: I think the work on this ticket is now really, "check that we are not unnecessarily copying PDU metadata for DECam raws in Gen3".  I suspect the behavior is already correct, but it's nontrivial to check so I'll leave that for the ticket.

It looks like we do read the primary header for DECam data as part of file ingest (it may even be read once by afw (because we don't yet know it's DECam data), and once by astropy as part of the per-detector read). I don't know if astropy supports INHERIT.

The raw formatter does not read the primary HDU, but it tries to read every HDU in order until it finds the correct one for the requested detector. I do see that DECam data does name the HDUs so I wonder if the raw formatter could instead calculate the detector name and then ask the FITS reader for that detector name explicitly.