Time-window logic between GSI and JEDI is inconsistent for sondes

[delippi]

Description

For sonde observations, GSI applies special logic in read_prepbufr.F90 that overrides the twindow specified in the convinfo table. Even when twindow = 1.5 hours (+/-5400 s), the 01z and 13z cycles will force late/early sondes into the window by resetting their time_drift to -1 hour and assimilate the full profile.

In practice, this means:

• All sonde obs older than the convinfo window are shifted to -3600 s
• They are still assimilated, even though they fall outside the nominal +/-5400 s
• A large spike appears at -3600 s in the GSI time distribution
• The same sonde profile is (likely) assimilated in two consecutive cycles (00z + 01z, and 12z + 13z)

JEDI does not implement this fallback logic, and simply applies the user-defined time window via MetaData/timeOffset. Thus, JEDI keeps only the obs that truly fall within +/-twindow, while GSI forcibly pulls older profiles back in.

Long-term question(s):

• Should RRFSv1 GSI be updated to respect the convinfo time window without forcing late/early soundings back in?
• Should JEDI eventually replicate this behavior, or should we align both systems to the simpler and more transparent JEDI time window?

[delippi]

This is the relevant code snippet:
https://github.com/NOAA-EMC/GSI/blob/13b754cc5edbf3fd276a75d9de91c9f354a534fc/src/gsi/read_prepbufr.F90#L1970-L1972

                    if(abs(time_drift) > ctwind(nc) .or. abs(time_drift) > twindin)then
                       time_drift=timex
                       if(abs(timex) > ctwind(nc) .or. abs(timex) > twindin) CYCLE LOOP_K_LEVS

This checks if the time_drift is outside the bounds set in convinfo table and then resets it to timex which is -1 h. This is why we see a huge spike at -1 h (-3600 s) in the figure:

I tested changing the code to:

                    if(abs(time_drift) > ctwind(nc) .or. abs(time_drift) > twindin)then
-                       time_drift=timex
+                      time_drift=time_drift
-                       if(abs(timex) > ctwind(nc) .or. abs(timex) > twindin) CYCLE LOOP_K_LEVS
+                      if(abs(time_drift) > ctwind(nc) .or. abs(time_drift) > twindin) CYCLE LOOP_K_LEVS

which resulted in matching obs counts (between GSI and JEDI; also using ext_sonde=.false.) and no peak at -3600s.

Edit:
For ext_sonde, see:
#495