[mkuemmel]

Checked it also on Raphael's dataset. And it does identify the bad apples.

[mkuemmel]

Perfect, also the check now happens earlier:

2025-09-26T19:42:03CEST BackgroundModel  INFO : Background for image: Masked(Masked(ProcessedImage(BufferedImage(/project/ls-mohr/users/martin.kuemmel/cluster_sepp_example/CORR-0268528-102.fits),ConstantImage<2048,4176>(0.070091))) with 0.0% weight image masked pixels;) with 38.0% weight image masked pixels; median: 0.0968899 rms: 4.06483!
2025-09-26T19:42:05CEST PsfPlugin ERROR : Failed loading a psf file: /project/ls-mohr/users/martin.kuemmel/cluster_sepp_example/CORR-0091740-102.psf
2025-09-26T19:42:05CEST SourceXtractor FATAL : Malformed variable PSF, coefficient matrices contains NANs
2025-09-26T19:42:05CEST Python::Interpreter  INFO : Python GIL acquired 63 times
INFO:    Cleaning up image.

[mkuemmel]

With only the good PSFEx files everything is golden:

bash-5.1$ tail euclid_slurm.usm-cl-bt01n2.11443608.err
2025-09-27T07:37:09CEST Progress  INFO : Elapsed: 10:29:39
2025-09-27T07:37:09CEST SourceXtractor  INFO : total 189544 sources detected
2025-09-27T07:37:09CEST ElementsProgram  INFO : ##########################################################
2025-09-27T07:37:09CEST ElementsProgram  INFO : #
2025-09-27T07:37:09CEST ElementsProgram  INFO : #  C++ program:  sourcextractor++ stops 
2025-09-27T07:37:09CEST ElementsProgram  INFO : #
2025-09-27T07:37:09CEST ElementsProgram  INFO : ##########################################################
2025-09-27T07:37:09CEST ElementsProgram  INFO : ##########################################################
2025-09-27T07:37:09CEST Python::Interpreter  INFO : Python GIL acquired 63 times
INFO:    Cleaning up image...

That was: 10 true cores, 20GB RAM and 10h run time.

[embray]

Unfortunately, as I've discovered while continuing to try to fix #649, this is not a complete fix. Even with the type punning modern versions of clang are clever enough to recognize this pattern. It "reasons", "x was a float, so even if we're punning it as an integer, because -ffast-math NaNs do not exist so this bit pattern should never be found in a float either". At least that appears to be what's happening because on my experiments with compiling with newer clang (21.1.0, from conda in my case) this check is completely compiled out on release builds.

One workaround is to compile with -fno-finite-math-only and do accept that Nan/Inf can exist and must be dealt with appropriately. I am also experimenting with some possible more targeted approaches.