John Swinbank, could you have a look? I've attached some before and after fix plots and have confirmed that the catalog is now computing the psf shapes at the source position rather than just the image average position for each source:

before:

In [10]: srcBefore["base_SdssShape_psf_xx"]

Out[10]: 

array([ 4.35643166,  4.35643166,  4.35643166, ...,  4.35643166, 4.35643166,  4.35643166])

after (for same visit/ccd combo);

In [12]: srcAfter["base_SdssShape_psf_xx"]

Out[12]: 

array([ 4.35188276,  4.35838516,  4.37397443, ...,  4.31416827, 4.31451534,  4.31364129])

Jenkins is happy: https://ci.lsst.codes/job/stack-os-matrix/label=centos-7,python=py2/19409/console

Plots look nice. Code change seems to be unambiguously the right thing. Well found.

But... (& please don't hate me for this) any chance of dropping in a test case to make sure this doesn't happen again? I've not looked at the SdssShape tests to see how straightforward it would be.

I don't hate you at all, and in fact already looked into it. It's not entirely straight-forward to test for this the way we're currently using testSdssShape.py to test the basic functionality as the test image built is almost as simplistic as it gets, only having one psf model for the entire image (so you'd get the same psf shape at any position of the image). I'll have a closer look to see if it can be adapted...

Ok, many thanks to Yusra AlSayyad for showing me how to create a variable psf model that could be applied to an exposure so that we could put in a unittest for this issue. The only hitch is that it requires use of meas_algorithms's PcaPsf, so the test can't live in meas_base. I've added tests/testSdssShapePsf.py to meas_algorithms. Below I show a printout for the measurements at 3 different source locations to ensure the psf is actually varying:

source position:  50.1 49.8

     psfTruth =  (ixx=3.89667634765, iyy=3.89667634765, ixy=-0.0)

sdssShape_psf =  (ixx=3.89668177791, iyy=3.89668177791, ixy=-0.0)

source position:  -11.6 -1.7

     psfTruth =  (ixx=2.97828641118, iyy=2.97828641118, ixy=-0.0)

sdssShape_psf =  (ixx=2.978288526, iyy=2.978288526, ixy=-0.0)

source position:  149.9 50.3

     psfTruth =  (ixx=8.99930898191, iyy=8.99930898191, ixy=0.0)

sdssShape_psf =  (ixx=8.99930656829, iyy=8.99930656829, ixy=0.0)

I also added a note to tests/testSdssShape.py in meas_base noting this.

Jenkins is down, but the test passes on my local build. Feel free to have a look when you have time and I'll let you know when I get the Jenkins run going.

Jenkins is happy: https://ci.lsst.codes/job/stack-os-matrix/label=centos-7,python=py2/19428/console

Added a fairly trivial comment to the PR on meas_algorithms. Otherwise, ok to merge.

By the way, the current standards say you should create your own PRs for review. See the discussion here. Not a big deal, obviously!

Happy holidays.

Thanks John Swinbank. I added the extra checks, reran Jenkins, and merged to master.