This could be considered a breaking change - not sure if this can just be merged?

I think it can be, as long as no implementation uses it yet. Which is covered largely by the discussion in gh-125 and in pola-rs/polars#10267 I believe.

However, to make extra sure, let's open PRs that remove the keyword where needed to implementers or ping the relevant authors. So far we covered:

• pandas
• Polars
• PyArrow

In addition, we know of:

• Modin (Cc @anmyachev as author of FEAT-#0000: enable interchange protocol for Series modin-project/modin#6361)
• Vaex (implementation at https://github.com/vaexio/vaex/blob/15245cf4332d4423ac58bd737aee27d911a1b252/packages/vaex-core/vaex/dataframe_protocol.py#L720)
• static-frame (implementation at https://github.com/static-frame/static-frame/blob/159039d4ec05001990393142cca7928ca04a011a/static_frame/core/protocol_dfi.py#L276)

They all have the keyword in their signatures, so even if it raises when a non-default value is passed in, it can't just be removed from the signature of __dataframe__ in the implementations, or it's going to break other libraries calling it with nan_as_null=nan_as_null.

So perhaps we change the description in the docs here first to "DO NOT USE", then remove it from the consumers calling __dataframe__, and only later on remove it from the signature of __dataframe__?