MDF IOAlg cannot run with recent simulation data (Sim10aU1)

When running a simple Moore configuration using the input files available at /daqarea1/fest/202110/mdf/30000000/ (in the online machines), an error occurs:

HLTControlFlowMgr                      INFO Starting loop on events
HLTControlFlowMgr                      INFO Timing started at: 16:34:18
HLTControlFlowMgr                      INFO Timing stopped at: 16:34:18
LHCb__MDF__IOAlg                      ERROR Bad magic pattern in Tell1 bank : Size:2304 Type:128:UNKNOWN Source:-9748 Vsn: 0 4c5a
LHCb__MDF__IOAlg                      ERROR Maximum number of errors ( 'ErrorMax':1) reached.
HLTControlFlowMgr                     FATAL Event failed in Node DeterministicPrescaler/DeterministicPrescaler
LHCb__MDF__IOAlg                      ERROR Bad magic pattern in Tell1 bank : Size:3328 Type:146:UNKNOWN Source:-2620 Vsn: 0 4c5a
LHCb__MDF__IOAlg                      ERROR Maximum number of errors ( 'ErrorMax':1) reached.
HLTControlFlowMgr                     FATAL Event failed in Node DeterministicPrescaler/DeterministicPrescaler#1
LHCb__MDF__IOAlg                      ERROR Bad magic pattern in Tell1 bank : Size:5376 Type:105:UNKNOWN Source:-22356 Vsn: 0 4c5a
LHCb__MDF__IOAlg                      ERROR Maximum number of errors ( 'ErrorMax':1) reached.
HLTControlFlowMgr                     FATAL Event failed in Node DeterministicPrescaler/DeterministicPrescaler#2
LHCb__MDF__IOAlg                      ERROR Bad magic pattern in Tell1 bank : Size:59392 Type:116:UNKNOWN Source:29832 Vsn: 1 4c5a
LHCb__MDF__IOAlg                      ERROR Maximum number of errors ( 'ErrorMax':1) reached.
HLTControlFlowMgr                     FATAL Event failed in Node DeterministicPrescaler/DeterministicPrescaler#3

It only happens when we set

options.use_iosvc = True
options.event_store = 'EvtStoreSvc'

A small reproducer is attached (reproducer.py).

Using mdfdump.py (which is fixed in !3290 (merged)), we confirmed that the magic patterns in the input from daqarea1 are correct. However, we noticed that the magic pattern obtained in DAQ/MDF/src/Buffer.cpp is always 19546 (0x4C5A) instead of 52171 (0xCBCB).

If we use a file from TestFileDB.py as input ('MiniBrunel_2018_MinBias_FTv4_MDF'), it runs fine regardless of the options above.

FYI @rmatev @sstahl @fsouzade