import numpy as np
import pandas as pd

vals = [1.5, 2.5]

pd.Series(vals, dtype=np.int64)  # <- float64
pd.Series(np.array(vals), dtype=np.int64)  # <- float64
pd.Series(np.array(vals, dtype=object), dtype=np.int64)  # <- raises
pd.Series(vals).astype(np.int64)  # <- int64

pd.Index(vals, dtype=np.int64)  # <- raises
pd.Index(np.array(vals), dtype=np.int64)  # <- float64
pd.Index(np.array(vals, dtype=object), dtype=np.int64)  # <- raises
pd.Index(vals).astype(np.int64)  # <- int64

xref #49372

When passing non-round float-like data and an integer dtype we have special logic

1. when a user asks for specific dtype, we should either return that dtype or raise
2. pd.Index behavior should match pd.Series behavior
3. pd.Series(values, dtype=dtype) should match pd.Series(values).astype(dtype)

I would be OK with either returning integer dtype or raising here. If we prioritize consistency with .astype behavior then this will depend on decisions in #45588.

Currently we have 13 tests that fail if we disable this behavior (9 of which are a single parametrized test) and they are all directly testing this behavior, so changing it wouldn't be world-breaking.