Interesting, Python allows this to drop through to concatenation even though both operators define normal addition (even though both operators define normal addition). I am not sure that is quite correct on the Python side, but I guess that doesn't really matter.

I can't really think of a solution right now. We may just have to delete that whole code path again, if it is accessible through the normal + operator.

We may just have to delete that whole code path again

Do you mean what was added in #16570?

Bah, this is all super awkward... Yes, I mean that PR, in the end I am not invested in keeping that PR around. I think that concat is weird, since returning NotImplemented should also guarantee Python's "incompatible operands TypeError", and that is something only we can fix.

But... At the same time pandas is sligthly breaking the "assymetry of binary operands" contract (which is very arcane!).

Basically __add__ and __radd__ have an inherent assymetry in that __radd__ has a bit of an obligation not to defer back (unless it is to a superclass).
Since you indicate to NumPy that it must defer from __add__ (using __array_priority__) you also get into the position that you are obliged to handle the operation! You are allowed to defer from your __add__, but not from __radd__ (although I am not sure how that distinction plays in cython).

To proof my point, try this:

arr = np.array([pd.Timedelta("10 days")])
pd.Timestamp("2021-01-01") + arr
# works: array([Timestamp('2021-01-11 00:00:00')], dtype=object)
arr + pd.Timestamp("2021-01-01")
# TypeError: unsupported operand type(s) for +: 'numpy.ndarray' and 'Timestamp'

In the first case it is OK that you defer to NumPy. NumPy realize its __radd__ is used and wraps the timestamp into an object array. In the second case, NumPy has no way of doing this (this is impossible in the way binary operators are designed in Python – I don't think it even depends on the C-API).

(Long story short, I won't complain about undoing the PR to get the old TypeError back, but I am not quite sure its actually the best outcome for either project.)

This happened to me because I forgot to convert a torch tensor to numpy. Basically in this scenario:

import torch
import numpy as np

a = np.array([1, 2,])
b = torch.tensor([1, 2,])
c = a + b

Error:

E           TypeError: Concatenation operation is not implemented for NumPy arrays, use np.concatenate() instead. Please do not rely on this error; it may not be given on all Python implementations.

This was quite confusing to debug because I used assignment to a slice where I've combined scatter on one dimension and bool mask on other dimension and LLMs would tell me the problem is with my indexing...