Salted Challenge Response HTTP Authentication Mechanism
draft-ietf-httpauth-scram-auth-15

-- Section 3 --

   It sends
   the username to the server, which retrieves the corresponding
   authentication information, i.e. a salt, StoredKey, ServerKey and the
   iteration count i.

As this is written, it's not clear whether the "corresponding authentication information" is "a salt" alone, or the whole list of things after.  I suggest this:

NEW
   It sends
   the username to the server, which retrieves the corresponding
   authentication information: a salt, a StoredKey, a ServerKey,
   and an iteration count (i).
END

Why does the paragraph end in a colon ("and sends a ClientProof to the server:")?

   To begin with, the SCRAM client is in possession of a username and
   password (*) (or a ClientKey/ServerKey, or SaltedPassword).

...and...

   (*) - Note that both the username and the password MUST be encoded in
   UTF-8 [RFC3629].

I suggest that splitting these up with a bunch of text in between isn't a good way to do it.  I think this would be better:

NEW
   To begin with, the SCRAM client is in possession of a username and
   password, both encoded in UTF-8 [RFC3629] (or a ClientKey/ServerKey,
   or SaltedPassword).
END

There's no pressing need to use the word "MUST" there.

   Informative Note: Implementors are encouraged to create test cases
   that use both username passwords with non-ASCII codepoints.  In
   particular, it is useful to test codepoints whose "Unicode
   Normalization Form C" and "Unicode Normalization Form KC" are
   different.

It would be good to include an informative reference to the Unicode Normalization Forms document here, so people can easily look up what NFC and NFKC are.  It might also be good to say "Normalization Form Canonical Composition (NFC)" and "Normalization Form Compatibility Composition (NFKC)", rather than to use the mixed versions that you have here.

-- Section 4 --

   from the IANA "Hash Function Textual Names" registry (see
   http://www.iana.org) .

Better, I think, to use the direct URL for the registry (which is fine with IANA):
http://www.iana.org/assignments/hash-function-text-names/

Some of the issues below would be discuss points for a standards
track draft but I think clarifying these kinds of corner-case issues
during an experimental deployment should be fine. I think it'd be
really good if you noted some of these issues and that they were
things to clarify as part of the experiment. The interleaving issue
being perhaps the most worthy of noting.

- abstract: the claims seem somewhat overbroad, I hoped those would
be justified in the text, but I didn't see that tbh. I'd recommend
toning that down some.

- intro, 1st para, typo: s/have had/has had/

- intro, 3rd para: s/adoptation/an adaptation/ and s/SCRAM data/ The
SCRAM data/ in the same para, and s/adds/adds a/ similarly

- intro, bullet1: where you say "is not sufficient" that can be true
of cleartext password schemes too, the difference here is that the
protocol data units must require a dictionary attack, which is good
but mostly neutral if good back-end practices are in place, but bad
back-end practices can be in place with SCRAM just as much as with
cleartext pwds, so I think the claim is not justified in the end
unless you add "depending on sound implemenation practices" or some
such.

- intro, bullets 2&3: these should really have a caveat that a
dictionary attack works and would invalidate the claims made

- 2.1: it is not clear to me how LDAP or RADIUS could be used as an
authentication DB for SCRAM, unless the cleartext pwd is stored
there, which would invalidate the claims made in the intro, or are
there standard ways in both protocols to store SCRAM verifiers that
are not cleartext passwords? (That might be the case, I just don't
know.)

- 2.2: "may not be included" would be better as "is optional" as the
former can be read ambiguously to mean "must not be included."

- 2.2: Definition of "Hi" - it is odd to have the "i" twice in the
formula.

- 2.2: definition of "Hi" - what is INT(1) and why is it there?  I
think you need to explain the former and should explain the latter.

- section 5: Hmm - I didn't realise that HTTP really allows for two
401 messages within the same authentication. It isn't clear to me
that a browser and server can always correctly maintain state for
that. Are you sure that they can? Can't two requests for protected
resources cross over in HTTP causing server confusion (if the realms
differ and passwords differ) or maybe even a condition that might
not represent the client's actual wishes? (That said, I've not gone
and looked if this kind of interleaving is a real problem or not.)

- 5, initial client msg: "a random, unique nonce attributes" is
ambiguous - is it one ("a random") or many ("attributes"). I think
you need to be precise.

- 5, just before 5.1: "might have to drop" is very vague, why?  If
the server auth is broken I think a "MUST drop" has to be correct.

- 5, is it ok for a client to increment the nonce-count by >1?  (I
assume it is as a request could get lost, albeit that'd be rare and
mostly down to proxy bugs I suspect.)

- section 8: "a successor to SHA-1" is a bit coy and now that Tony
has defined one I think you can mention the sha256 based scheme.

- Thanks for handling the secdir review [1] and I appreciate that
the author has pinged the reviewer on that. It'd maybe be no harm to
re-ping if the reviewer still hasn't had a chance to respond. 

  [1] https://www.ietf.org/mail-archive/web/secdir/current/msg06254.html