Opted for simple solution: extending the bbox by a cell size width. This fixes a few things:
1) Matching kernels are spatially varying, and now the spatial model has data points beyond the image to fit. This means that the edges of the image are being fit properly rather than extrapolated.
2) Cell Sets cannot be partial. In 1D for example, this means if cellSize = 128pix, and the image is 255pix, there's one data point at 64pix. With this tickets update, there will be a data point at -64, 64 and 192.
3) Narrow slivers of calexp can now have sufficient data to fit an accurate matching kernel.
Note: that this bbox extension is IN ADDITION to WarpAndPsfMatchTask (pipe_tasks) bbox extension of the warped image. While that could be also improved, it needs to stay because otherwise you get border of NaNs after psfMatching.