AMGX's results fail to converge in my case

Hello everyone！
I am trying to use AMGX as the solver of my FEM program. I have used FGMRES_AGGREGATION.json and other configuration files, but none of them can converge. I'm sure my equation has a solution and I can get the correct result using the CG method (config_version=2, solver(cg)=CG, cg:preconditioner=NOSOLVER, cg:max_iters=1000, cg:tolerance=1e-6, cg :monitor_residual=1, cg:print_solve_stats=1).
What am I doing wrong? Or is the AMGX some kind of limitation that I'm not aware of?

Below is my result using FGMRES_AGGREGATION.json：
AMGX version 2.4.0
Built on Nov 22 2022, 09:51:51
Compiled with CUDA Runtime 10.2, using CUDA driver 11.0
AMG Grid:
Number of Levels: 5
LVL ROWS NNZ SPRSTY Mem (GB)
--------------------------------------------------------------
0(D) 3350 64708 0.00577 0.000822
1(D) 1605 26691 0.0104 0.000668
2(D) 769 12377 0.0209 0.000311
3(D) 367 5019 0.0373 0.000129
4(D) 176 2284 0.0737 5.63e-05
--------------------------------------------------------------
Grid Complexity: 1.87075
Operator Complexity: 1.71662
Total Memory Usage: 0.00198548 GB
--------------------------------------------------------------
iter Mem Usage (GB) residual rate
--------------------------------------------------------------
Ini 1.4845 3.162278e+04
.
.
.
989 1.4845 2.439992e+04 1.0000
990 1.4845 2.439992e+04 1.0000
991 1.4845 2.439992e+04 1.0000
992 1.4845 2.439992e+04 1.0000
993 1.4845 2.439992e+04 1.0000
994 1.4845 2.439992e+04 1.0000
995 1.4845 2.439992e+04 1.0000
996 1.4845 2.439992e+04 1.0000
997 1.4845 2.439992e+04 1.0000
998 1.4845 2.439992e+04 1.0000
999 1.4845 2.439992e+04 1.0000
--------------------------------------------------------------
Total Iterations: 1000
Avg Convergence Rate: 0.9997
Final Residual: 2.439992e+04
Total Reduction in Residual: 7.715932e-01
Maximum Memory Usage: 1.484 GB
--------------------------------------------------------------
Below is my test data
test.zip

Start with the attached test.zip and the FGMRES_AGGREGATION.json configuration, then compare its solver settings and output with the working CG configuration shown in the issue. Reproduce the reported residual behavior and determine whether the result comes from the input/configuration or an AMGX limitation; document the confirmed cause and a reproducible resolution.