It's looking like the interface to DM-29732 could be a lot cleaner if e.g. a master bias was saved with its own Detector.  That's because in order to get an amp subimage, one has to provide a Detector reflecting the full bias image to the code extracting the subimage, as well as an amp describing the subimage to extract, for two reasons:

• we need an existing Detector to make a single-amp Detector to return with the subimage (for all of the detector-wide information that doesn't change when it's restricted to one amp);
• it's possible for bias (and maybe some niche derived-from-raw datasets) to be trimmed or untrimmed, and looking at Detector.getBBox() (always trimmed) vs. the on-disk image bbox seems to be the best way to figure that out.

But right now it seems pretty difficult to get the appropriate Detector for a bias (even absent that question about whether it's trimmed or not):

butler.get("camera", ...)[detectorId] 

will get you the wrong amplifiers for LSST (reflecting the unassembled on-disk raws),

butler.get("bias", ...).getDetector()

# and

butler.get("bias.detector, ...)

are None,

and

 butler.get("raw.detector", ...)

doesn't work at all.  I'm going to fix the latter on DM-29370 (already on the branches), because it's something we should have been supporting before, and so we'll at least have something.

But that's an awkward way to get a detector for some other image, and for LATISS (at least), you can still get something that doesn't match an untrimmed bias because of the raw overscan patching that can happen - but I think that shouldn't be a problem in practice.

Anyhow, the easy way to fix this is to just propagate a Detector through the CPP combine tasks and save it with the Exposure.