Hi @OTAKU ,
I cannot confirm a supported reading contract covering your build 26200.9457 from the public material I checked. The TOKEN_INFORMATION_CLASS documentation still specifies DWORD. In Microsoft's win32metadata discussion, Jevan Saks states that the documentation should describe a BOOLEAN with a DWORD-aligned pointer, agrees that only the first byte is valid when ReturnLength == 1 with the remaining bytes outside the result, and treats any deviation as a documentation problem. He also expects the behavior to be consistent across versions, but that is an expectation rather than a published architecture or servicing scope covering your target.
My native C tests ran on a separate machine reporting build 26300.9457 with advapi32 10.0.26100.9278 and KernelBase 10.0.26100.9444. With an aligned four-byte buffer, class 21 succeeded with ReturnLength = 1 and only offset 0 changed; offsets 1 through 3 retained two different caller-supplied fill patterns. The size query failed with error 122 and also reported 1, while classes documented as DWORD, such as TokenVirtualizationAllowed and TokenUIAccess, reported 4 and changed all four bytes. Passing capacity 1 on a larger guarded allocation also succeeded when aligned; offsets of 1, 2 or 3 bytes failed with error 998 (ERROR_NOACCESS). These observations do not establish your target's supported minimum capacity or the exact physical write extent.
GetTokenInformation defines ReturnLength as the required buffer size, so that value alone does not resolve the output-type discrepancy. The documented zero/nonzero meaning is whether the token has ever been filtered. Under the BOOLEAN interpretation, value != 0 preserves that meaning; requiring value == 1 would be narrower. I observed only 0 and 1. Tokens created through CreateRestrictedToken with DISABLE_MAX_PRIVILEGE, LUA_TOKEN, or a deny-only SID returned 1 while IsTokenRestricted returned false, so class 21 is not the same check as IsTokenRestricted. A UAC-filtered linked token was not available on that machine, so I did not test that case. I cannot confirm that Windows emits other nonzero values or infer the byte value from your saved invocation.
I understand that this is not the clear-cut answer you were looking for, especially when your implementation is already stopping at this check and the public material only gets you partway there. What I have shared here comes from my own testing and the public documentation I could find, and that is honestly as far as I can take it from the forum side. For the written applicability statement you need, Windows Developer Support is the right place to ask; select the product that matches your scenario. Microsoft lists technical support there as paid incident-based support requiring a support contract. Include your existing request and ask for review by the responsible token API team, with a written conclusion covering valid bytes, capacity, alignment, nonzero semantics, and the supported architecture/servicing scope containing your target. That conclusion is the outstanding clarification needed before deciding whether to change your current decoding rule.
I appreciate how carefully you have scoped and documented this question; it made the review much easier, and I hope the information above helps you move forward with your assessment. If you find it useful, you are welcome to follow this guidance so others with a similar question can find it more easily.
Thank you.