Hi Pedro,
I created and run Windows NT service in almost all Visual Age Versions (Envy OTI 3, VAST 4 through 11 or 12) until just recently (1993 -2021) in several different constellations, environments and applications.
Even change of the bitness (16->32->64) (except in one case) caused problems.
I never observed a lock in any production version of a service (packages without a graphical user interface).
The evaluation versions block execution of the development environment after a given time period, but instantly at execution start, not showing the behaviour of your description. In particular the Smalltalk VM stops execution in its own initialisation, being run out of the test period.
Look at other potential reasons, internal (Smalltalk code Q&A) or external (access rights, DLLs, DB, ...) resources causing the lock.
Try to obtain external insights using tools like process explorer under Windows to gain more information: what is locked or contributing to the lost execution time. [for example, recently we found a memory leak accessing a DLL, causing tiny incremental memory loss, production code already operating several years ago].
Is there anything else causing timeouts and retries?
And consider all envolved (other) changes along the version swap.
So switching from 8.5 to 9.2.2 implied a generation exchange of the VM as I remember.
Did you properly replace the whole installation of the server runtime support?
It is not sufficent to just exchange the image alone (except in some tiny increments in rare cases, e.g. 6.3 -> 6.4).
Usually the whole server software is to be installed on a different place than before, so have an eye on all pathnames involved.
I did observe locks of an application in a server environment running several instances with GUIs in parallel - when one image crashed (due to other reasons) it was not possible to start any other image until the whole Windows VM session was closed and restarted (not only a new login, but the whole server has to be restarted).
We did not track that case further down. It seemed to be related to GUI logic of the login of the Smalltalk application, which was shared over all instances until the whole server was shut down (including an ODBC connection to be established).
This behaviour could not be reproduced in a client installation - the remote execution of the GUI was an essential part in this lock.
-