Fix IndexError when unpacking empty tuple in type alias

[bxff]

Fixes #20913

mypy crashes with IndexError when using Unpack[tuple[()]] (or just omitting the TypeVarTuple args entirely) in a PEP 695 type alias:

class C[*Ts]:
    pass

type T[T, *Ts] = C[*Ts]

x: T[bool, Unpack[tuple[()]]]  # IndexError
y: T[bool]                      # also crashes

The crash is in visit_instance() — when expanding builtins.tuple args that reduce to nothing, args[0] is accessed on an empty list.

The fix handles this in two places:

1. visit_instance: when builtins.tuple ends up with empty args after expansion, return TupleType([], ...) (the empty fixed-length tuple) instead of crashing
2. expand_type_list_with_unpack / expand_type_tuple_with_unpack: inline Unpack[TupleType] results into their items, so Unpack[tuple[()]] correctly expands to nothing

Both T[bool] and T[bool, Unpack[tuple[()]]] now correctly resolve to C[()].

[A5rocks]

Can you add a reveal type to make sure the type is being preserved?

[bxff]

Done. reveal_type(x) -> __main__.C[Unpack[builtins.tuple]]

[A5rocks]

Ok this is the wrong type.

[bxff]

Agreed, but this is pre-existing, on unpatched master (without this fix), C[Unpack[tuple[()]]] already reveals identically to unparameterized C:

# on master, no fix applied
x: C[Unpack[tuple[()]]]
reveal_type(x)   # C[Unpack[builtins.tuple[Any, ...]]]

y: C
reveal_type(y)   # C[Unpack[builtins.tuple[Any, ...]]]

This PR only prevents the IndexError crash; the wrong type for empty tuple unpacks is a separate issue.

[A5rocks]

I'm not sure that's identical behavior (note that the messages there have parameters provided to tuple). IIRC the issue you're showing is that we don't pass empty_index (or whatever the name is)... but in this case I don't think that's relevant, because if you unpack something empty that should be like x: T[bool] which currently has the reveal_type output.

Basically, I think what you show and this PR are different code paths (based on what I remember), so I think that behavior is irrelevant...? I'd be fine if this is a followup tbh, as long as you poke at it a bit and make sure it's not a 1 line change or something wrong with your crash fix.

[bxff]

Ahh I see my mistake, you're right, I was comparing against the wrong code path. T[bool, Unpack[tuple[()]]] should just be T[bool] which gives C[()]. Fixed and updated the test.

[A5rocks]

Gah, no, I misread what T was. My bad! Interesting that this also crashes:

class C[*Ts]:
    pass

type T[T, *Ts] = C[*Ts]

x: T[bool]

reveal_type(x)

So no Unpack[tuple[()]] necessary. Given that, I'm a bit surprised about the code modifications you made -- what's the tuple instance without .args coming from? (unless that's a different error!)

[bxff]

Same error actually. T[bool] hits the same args[0] IndexError at expandtype.py:228. Tested both on 3.12 with the fix:

x: T[bool]                        -> C[()]
y: T[bool, Unpack[tuple[()]]]     -> C[()]

I'll add T[bool] as a test case too.

[A5rocks]

Hmm, so in args[0] it's synthesized based on the args from T, I assume? Interesting.... Sorry for all the comments! I'm not comfortable enough with this part of mypy to know if this is the right approach.

[bxff]

No worries, appreciate the review! The T[bool] catch was really helpful.

[A5rocks]

I'm not completely comfortable with this, but I'm not familiar with this part of mypy so that doesn't mean anything :/

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅