Wrong inference for `str(const)`

[jkmnt]

These return Const(""), which is really misleading (even an Uninferable is better):

code = """
    s = str("abc") #@
    s = str(1) #@
    s = str(None) #@
    s = str(2 + 2) #@
    """

Looking at brain_builtin_inference.py, the def infer_str() function is hardcoded to always return Const("").
Next to it is the def infer_int(), which at least tries to examine int() arguments.

I'll submit a small PR fixing it at least for const argument.

[jkmnt]

On a side note: looking closer to the builtins inference implementation, there is common pattern of returning str(""), int(0), [], {} etc if argument inference went bad. Is it by design ? Wouldn't it be better to raise UseInferenceDefault ? Then the callers are likely to fallback to the Instance(builtin.str), Instance(builtin.int), etc, which is the correct answer IMHO.

[Pierre-Sassoulas]

I can't find a discussion argumented in detail about this in a short time but most of the time it's advantageous to return "it's a string" even if we don't know the content of it. Some check in pylint stop abruptly on inference fail, and most of the time "string" is good enough. The fallback would work better and be smarter but in pylint most of the time we simply call safe_infer and do not consider the various inference fail in the check (we probably should).

[DanielNoord]

I agree with @Pierre-Sassoulas

In hindsight it would probably be better to do so, but I'm afraid it will cause a lot of churn for pylint and other users of astroid. So not sure if it is worth the effort to change it.

[Pierre-Sassoulas]

pylint is tested well enough automatically that an agentic AI could be remarquably efficient fixing this in pylint. But it would still affect plugins. Sounds like an issue that need investigation. (Maybe a new pylint-speficic checker/plugin to suggest catching inference error explicitely when inferreing + deprecation would make this less painful? There was some exploration on inference error catching earlier but using the generic safe_infer wasn't flagged as bad pragtice at the time)