BUG: Index.insert should coerce to the appropriate type

[jreback]

xref to #14535 (PR to fix, but stale)

# NG, must be float dtype
pd.Index([1]).insert(1, 1.0)
# Int64Index([1, 1], dtype='int64')

# NG, must be object dtype
pd.DatetimeIndex(['2011-01-01']).insert(1, 1)
# TypeError: cannot insert DatetimeIndex with incompatible label

Expected

pd.Index([1]).insert(1, 1.0)
# Float64Index([1.0, 1.0], dtype='float64')

pd.DatetimeIndex(['2011-01-01']).insert(1, 1)
# Index([2011-01-01 00:00:00, 1], dtype='object')

[jreback]

cc @gfyoung @toobaz if you have interest

[gfyoung]

How functional is the patch by @sinhrks ? Seems we would just need to pull down his branch and then just rebase for merging.

[jreback]

yes it's prob almost there

[jbrockmendel]

OK to just fix or need a deprecation cycle?

• IntervalIndex raises on mismatch
• PeriodIndex casts to object on mismatch (not even trying to cast strings, which is inconsistent with other setitem-like methods)
• DTI/TDI cast to object when they see a str (not even trying to cast), otherwise raises on mismatch
• CategoricalIndex raises on mismatch
• NumericIndex if anything try too hard to cast, e.g. inserting True will cast to 1, which I think it should not.

[jbrockmendel]

after #40096 all thats left is IntervalIndex (and some tuple cases)

[jbrockmendel]

IntervalIndex now does this too, though not sure when that change was made. Closing as complete.