Put original string instead of foldedCase

[lyokha]

The foldedCase is derivative of the original string, while the latter
is unique. So, putting foldedCase instead of original string leads to
loss of data, while getting from the original string must always restore
CI object correctly as it has been built from the original string.

Probably it should also fix weird cases like in issue #7.

My own case was restoring serialized ResponseHeaders from
Network.HTTP.Types. Say, putting Custom-Header and restoring it later
makes it custom-header (when I take it with original), while when
putting original string it restores correctly as Custom-Header (when I
take it with original).

[phadej]

IIRC my reasoning for using foldedCase is because I wanted to keep a == b -> encode a == encode b property. If your type is really CI a, you shouldn't care about this difference.

Specifically, HTTP headers are case insensitive, thus CI is their type.

[lyokha]

Yes, I see the reason. To be precise, equation a == b -> encode a == encode b always holds as far as encoding algorithm must be deterministic, did you mean a == b <-> encode a == encode b? If so, then this invariant does not hold in the current implementation as well when original data is known: the data simply gets lost while being serialized (i.e. encode a == encode b while originally a /= b). The invariant only holds when original data is not known.

Nevertheless, in my case I wanted to forward HTTP headers from a backend to the client, so I wanted to keep them equal (yes, I know that they are case-agnostic, however bad client implementations sometimes matter).

[phadej]

To be precise, equation a == b -> encode a == encode b always holds as far as encoding algorithm

>>> CI.mk "FOO" == CI.mk "foo"
True

You confuse == for the structural equality. It isn't.

[lyokha]

I worked this around by putting original headers in the encoded value and then re-applying mk to them when needed. So, the PR can be closed.