Strange constraints generated by GUT inference for type variable uses

[AndrewShf]

Testcase:

import universe.qual.*;

class Demo<T extends Box> {
    T f;
    void test() {
        f.v = new Object();
    }
}

class Box {
    Object v;
}

Using debug solver, we can know how varslot is inserted and constraints are generated.

class Demo<@VarAnnot(8) T extends @VarAnnot(9) Box> {
    @VarAnnot(7)
    T f;
    void test(@VarAnnot(15) Demo<@VarAnnot(16) T> this) {
        f.v = new @VarAnnot(20) Object();
    }
}

class Box {
    @VarAnnot(22)
    Object v;
}

Several constraints are questionable
18 = ( 15 ▷ 10: ( 7 | 9 ) )
19 = ( 15 ▷ 11: ( 7 | 8 ) )
20 <: 26
22 == 5 @ Rep weight = 80
23 = ( 10: ( 7 | 9 ) ▷ 22 )
24 != 3 @ Lost
24 = ( 10: ( 7 | 9 ) ▷ 23 )
25[ ↧ 24 ] == 20
26 = ( 10: ( 7 | 9 ) ▷ 25[ ↧ 24 ] )

this.f has viewpoint adaptation of the lower and upper bound, which is encoded by 19 = ( 15 ▷ 11: ( 7 | 8 ) ) and 18 = ( 15 ▷ 10: ( 7 | 9 ) ) respectively.

However, if 23 = ( 10: ( 7 | 9 ) ▷ 22 ) represents f.v, it's an incomplete viewpoint adaptation. it should use varslot 18 computed earlier for viewpoint adaptation, rather than varslot 10.

didn't understand why reuse varslot 10 to viewpoint adapt 23 again.

varslot 25 is the refinement variable, but why use that for viewpoint adaptation for varslot 26? Then there is a subtype constraint from 20 to 26 (which is likely to the the this.v = new object() assignment).