Towards the goal of removing Json from stdlib, this change replaces
usage of TotallyNotJson in test_gen/gen_abilities with a simple usable
inline implementation of Encoder/DecoderFormatting.
Similarly, the use of TotallyNotJson in test_reporting is not necessary
at all and is replaced with a Decoder that wouldn't actually work, but
which does compile.
A "late specialization" of a type is an ability specialization that
is not visible or needed until after type-specialization; i.e. during
monomorphization.
The `Inspect.toInspector` ability is special-cased for opaques that do
not claim or explicitly implement `Inspect`. In such cases, they are
treated as structural types, and given the immediate specialization of
`Inpect.inspectOpaque`.
However, prior to this commit, that special-casing would only be applied
during early specialiation (i.e. specializations visible during
generalized type inference). This commit applies the special case to
late specialization as well - the specialization decision for an opaque
type is always the specialization of the opaque type in the late case,
but now, when we go to look up the ambient lambda set of the
specialization, if it does not exist and corresponds to
`Inspect.toInspector`, we fall back to the immediate.
One concern I have here is that in a case like
```
Op := {}
x =
dbg (@Op {})
```
the specialization of `Inspect.toInspector` for `Op` should be known
early. Indeed, the program
```
Op := {}
x =
Inspect.toInspector (@Op {}) |> Inspect.apply (Inspect.init {}) |> Inspect.toDbgStr
```
Compiles fine without this change. This makes me suspect there is an
issue with the implementation of `dbg`'s desugaring. If possible, this
should be addressed sooner rather than later.
Closes#6127
