fermipy v1.4.1 (and possibly dev version) results are not stable

[HealthyPear]

Describe the bug

This is a bug I found while testing issue #566 after the new release.

Same exact dataset, same notebook run multiple times, once I see SED points as they should be
next time (or 3rd time) all flux errors are gone and sed["ts"] is nan in all bins

To Reproduce

Same dataset as per PG 1553 tutorial but 4FGL-DR4 version, namely:

Query Parameters (L2603111134438994BE7968)
Equatorial coordinates (degrees) | RA = 238.929, Dec = 11.1901
Time range (MET) | (239557417, 256970880)
Time range (Gregorian) | (2008-08-04 15:43:36, 2009-02-22 04:47:58)
Energy range (MeV) | (100, 300000)
Search radius (degrees) | 30

data/gll_iem_v07.fits
data/gll_psc_v35.fit
data/iso_P8R3_SOURCE_V3_v1.txt
data/L2603111134438994BE7968_PH00.fits
data/L2603111134438994BE7968_PH01.fits
data/L2603111134438994BE7968_SC00.fits

this is my config.yaml

data:
  evfile: data/ft1.txt
  scfile: data/L2603111134438994BE7968_SC00.fits

binning:
  roiwidth: 10.0
  binsz: 0.1
  binsperdec: 8

selection:
  emin: 100
  emax: 300000
  zmax: 90
  evclass: 128
  evtype: 3
  target: "4FGL J1555.7+1111"
  tmin: 239557414
  tmax: 428903014

gtlike:
  edisp: True
  irfs: "P8R3_SOURCE_V3"
  edisp_disable: ["isodiff", "galdiff"]

model:
  src_roiwidth: 10.0
  galdiff: "/Users/michele/Software/tests/fermipy/pg1553/data/gll_iem_v07.fits"
  isodiff: "/Users/michele/Software/tests/fermipy/pg1553/data/iso_P8R3_SOURCE_V3_v1.txt"
  catalogs:
    - "/Users/michele/Software/tests/fermipy/pg1553/data/gll_psc_v35.fit"

fileio:
  outdir: results

my analysis is a simplified version of the tutorial notebook

from fermipy.gtanalysis import GTAnalysis
gta = GTAnalysis('config.yaml',logging={'verbosity': 3})

gta.setup()

gta.free_sources(distance=3.0,pars='norm')

gta.free_source('galdiff')
gta.free_source('isodiff')

target_source = '4FGL J1555.7+1111'

gta.free_source(target_source)

fit_results = gta.fit()

gta.write_roi('fit0',make_plots=True)

sed = gta.sed(target_source)

print(sed["ts"])

import numpy as np
import matplotlib.pyplot as plt

c = np.load('results/fit0.npy', allow_pickle=True).flat[0]
E = np.array(c['sources'][target_source]['model_flux']['energies'])
dnde = np.array(c['sources'][target_source]['model_flux']['dnde'])
dnde_hi = np.array(c['sources'][target_source]['model_flux']['dnde_hi'])
dnde_lo = np.array(c['sources'][target_source]['model_flux']['dnde_lo'])

plt.clf()
plt.loglog(E, (E**2)*dnde, 'k--')
plt.loglog(E, (E**2)*dnde_hi, 'k')
plt.loglog(E, (E**2)*dnde_lo, 'k')
plt.errorbar(sed['e_ctr'][:-2],
             sed['e2dnde'][:-2],
             yerr=sed['e2dnde_err'][:-2], fmt ='o')
plt.errorbar(np.array(sed['e_ctr'][-2:]),
         sed['e2dnde_ul95'][-2:], yerr=0.2*sed['e2dnde_ul95'][-2:],
             fmt='o', uplims=True)
plt.xlabel('E [MeV]')
plt.ylabel(r'E$^{2}$ dN/dE [MeV cm$^{-2}$ s$^{-1}$]')
plt.show()

Expected behavior

I should always see the first SED (if I do not change data and analysis code in any way, should I get exactly the same plot? not sure if the fit starts internally from a random seed, but it should be more or less the same no? I hope!)

but instead at the 2 or 3rd time I see

Log files

This log file should contain multiple runs of the notebook, the last one shows the plot with only points
and no error bars

fermipy.log

This is a sped-up gif from a screen recording that shows what I describe

Environment (please complete the following information):

• OS: (e.g. MacOS Monterey version 12.3.1)

macos 15.7.4

• Output of uname -a

Darwin host-10-1-3-145.byod.mppmu.mpg.de 24.6.0 Darwin Kernel Version 24.6.0: Mon Jan 19 21:59:23 PST 2026; root:xnu-11417.140.69.708.3~1/RELEASE_ARM64_T6030 arm64

• fermipy, fermitools, and fermitools-data versions (for exameple, fermipy 1.0.1, fermitools 2.0.8)

fermipy version 1.4.1
ScienceTools version 2.5.1
Additional context
Add any other context about the problem here.