Infeasible model after consecutive invocation

We have the following workflow implemented in Java with Xpress 9.2.2 8.14.5:

1. Build BCL MIP model

2. Fix subset of variables (xprbVar.fix(.)). This simplified model is referred to as model A

3. Load model into optimizer (xprb.loadMat())

4. Solve model A (xprsProb.mipOptimize();) 

5. Unfix variables from step 2 (xprbVar.setLb(.), xprbVar.setUb(.)), thereby obtaining the original model. We refer to this model as model B

6. Load model into optimizer (xprb.loadMat())

7. Solve model B. If model A produced a feasible solution, we use this as a warmstart for B (xprb.loadMIPSol(mipSolution.x, true)). 

When executing this workflow, we ran into the following issue.

For a particular instance, model A is infeasible due to the variable fixing, but a solution exists to model B. When we execute the above workflow, this is what happens:

1. Model A returns infeasible (as expected)

2. Model B returns infeasible (not expected!):

Minimizing MILP LTLN using up to 8 threads and up to 15GB memory, with these control settings:
MAXTIME = -3600
MIPLOG = 3
THREADS = -1
MIPRELSTOP = .05
Original problem has:
      8258 rows         5407 cols        32026 elements      5403 globals
       850 delrows
Presolved problem has:
      3111 rows         3446 cols        14066 elements      3446 globals
      2642 delrows
LP relaxation tightened
Presolve finished in 0 seconds
Heap usage: 4527KB (peak 8518KB, 8219KB system)
 
Coefficient range                    original                 solved        
  Coefficients   [min,max] : [ 2.38e-04,  3.38e+05] / [ 5.95e-05,  2.00e+00]
  RHS and bounds [min,max] : [ 1.00e+00,  3.36e+04] / [ 4.05e-02,  6.20e+01]
  Objective      [min,max] : [ 1.00e+00,  1.00e+00] / [ 2.50e-01,  5.06e+03]
Autoscaling applied standard scaling
 
Will try to keep branch and bound tree memory usage below 4.4GB
Primal starts after 0 seconds
Crash basis containing 558 structural columns created
 
   Its         Obj Value      S   Ninf  Nneg        Sum Inf  Time
     0       66090.81967      p      9    30       7.882812     0
    19       66123.04311      p      5     0       5.000000     0
Problem is infeasible
 *** Search completed ***

Since this code is part of a large application, it's quite a challenge to create a small reproducible example. I however tried the following:

- when exporting model B after it is deemed infeasible, and running the exported lp or mps file through the interactive optimizer, the model solves correctly as expected.

- when searching for an IIS after model B was deemed infeasible, it doesn't seem to find an IIS:

General IIS
 
Problem is MIP feasible
 *** Search unfinished ***    Time:     0 Nodes:          1

- disabling the presolve (xprs.setIntControl(XPRS.PRESOLVE,0) seems to resolve the issue, i.e. model A is still infeasible as expected, but model B now correctly produces a solution.

 My suspicion is that the presolve eliminates some variables when solving model A. Then, after unfixing some variables and invoking solve again (model B), those variables that were eliminated by the presolve during the first invocation are still absent. This would be strange though, since xprb.loadMat() is invoked before every call to mipOptimize to ensure that the new updated variable domains are taken into account.

 Any suggestions on how to further troubleshoot this issue? Ideally I would prefer not to disable the presolve.