Hello,

ltlsynt 2.16.2 calls this formula unrealizable with its default options:

    ltlsynt -f '(!i0 & F(i1 & o3)) -> (i1 & o3 & G!i0)' --outs=o3 --realizability
    UNREALIZABLE

It is realizable: a controller that keeps o3 false makes F(i1 & o3) false, so the implication
holds. ltlsynt agrees once decomposition is off, and so does Strix:

    ltlsynt -f '(!i0 & F(i1 & o3)) -> (i1 & o3 & G!i0)' --outs=o3 --decompose=no --realizability
    REALIZABLE
    strix -f '(!i0 & F(i1 & o3)) -> (i1 & o3 & G!i0)' --ins=i0,i1 --outs=o3 -r
    REALIZABLE

With --verbose, ltlsynt splits the formula into two subformulas:

    (!i0 & F(i1 & o3)) -> (i1 & o3)
    !i0 -> G!i0

The second one has lost the assumption F(i1 & o3), which mentions the output o3. On its own it is
unrealizable (the environment sets i0 false, then true), so the whole formula is reported
unrealizable. It looks as if an assumption that mentions outputs is dropped from a subformula
whose outputs it does not share. The system can falsify such an assumption, so dropping it makes
the subformula harder to realize, not easier.

The formula was found by random testing (a random TLSF specification, reduced with ltlgrind while
ltlsynt with and without --decompose=no disagree). 

If you want, I could ask Claude Code to provide a patch.

Best regards,
Tom van Dijk