Internationalization is not a simple task, even if the consideration is limited to "merely" the textual aspect [1]. Languages turn out to be incredibly diverse in their writing systems, so software that tries to support all writing systems equally well ends up running into several problems that admit no general solution. Unfortunately, I am ill-placed to be able to offer personal experience with internationalization concerns [2], so some of the information I give may well be wrong.
A word of caution: this post is rather long, even by my standards, since the problems of internationalization are legion. To help keep this post from being even longer, I'm going to assume passing familiarity with terms like ASCII, Unicode, and UTF-8.
The first issue I'll talk about is Unicode normalization, and it's an issue caused largely by Unicode itself. Unicode has two ways of making accented characters: precomposed characters (such as U+00F1, ñ) or a character followed by a combining character (U+006E, n, followed by U+0303, ◌̃). The display of both is the same: ñ versus ñ (read the HTML), and no one would disagree that the share the meaning. To let software detect that they are the same, Unicode prescribes four algorithms to normalize them. These four algorithms are defined on two axes: whether to prefer composed characters (like U+00F1) or prefer decomposed characters (U+006E U+0303), and whether to normalize by canonical equivalence (noting that, for example, U+212A Kelvin sign is equivalent to the Latin majuscule K) or by compatibility (e.g., superscript 2 to a regular 2).
Another issue is one that mostly affects display. Western European languages all use a left-to-right, top-to-bottom writing order. This isn't universal: Semitic languages like Hebrew or Arabic use right-to-left, top-to-bottom; Japanese and Chinese prefer a top-to-bottom, right-to-left order (although it is sometimes written left-to-right, top-to-bottom). It thus becomes an issue as to the proper order to store these languages using different writing orders in the actual text, although I believe the practice of always storing text in "start-to-finish" order, and reversing it for display, is nearly universal.
Now, both of those issues mentioned so far are minor in the grand scheme of things, in that you can ignore them and they will still probably work properly almost all of the time. Most text that is exposed to the web is already normalized to the same format, and web browsers have gotten away with not normalizing CSS or HTML identifiers with only theoretical objections raised. All of the other issues I'm going to discuss are things that cause problems and illustrate why properly internationalizing email is hard.
Another historical mistake of Unicode is one that we will likely be stuck with for decades, and I need to go into some history first. The first Unicode standard dates from 1991, and its original goal then was to collect all of the characters needed for modern transmission, which was judged to need only a 16-bit set of characters. Unfortunately, the needs of ideographic-centric Chinese, Japanese, and Korean writing systems, particularly rare family names, turns out to rather fill up that space. Thus, in 1996, Unicode was changed to permit more characters: 17 planes of 65,536 characters each, of which the original set was termed the "Basic Multilingual Plane" or BMP for short. Systems that chose to adopt Unicode in those intervening 5 years often adopted a 16-bit character model as their standard internal format, so as to keep the benefits of fixed-width character encodings. However, with the change to a larger format, their fixed-width character encoding is no longer fixed-width.
This issue plagues anybody who works with systems that considered internationalization in that unfortunate window, which notably includes prominent programming languages like C#, Java, and JavaScript. Many cross-platform C and C++ programs implicitly require UTF-16 due to its pervasive inclusion into the Windows operating system and common internationalization libraries [3]. Unsurprisingly, non-BMP characters tend to quickly run into all sorts of hangups by unaware code. For example, right now, it is possible to coax Thunderbird to render these characters unusable in, say, your subject string if the subject is just right, and I suspect similar bugs exist in a majority of email applications [4].
For all of the flaws of Unicode [5], there is a tacit agreement that UTF-8 should be the character set to use for anyone not burdened by legacy concerns. Unfortunately, email is burdened by legacy concerns, and the use of 8-bit characters in headers that are not UTF-8 is more prevalent than it ought to be, RFC 6532 notwithstanding. In any case, email explicitly provides for handling a wide variety of alternative character sets without saying which ones should be supported. The official list [6] contains about 200 of them (including the UNKNOWN-8BIT character set), but not all of them see widespread use. In practice, the ones that definitely need to be supported are the ISO 8859-* and ISO 2022-* charsets, the EUC-* charsets, Windows-* charsets, GB18030, GBK, Shift-JIS, KOI8-{R,U}, Big5, and of course UTF-8. There are two other major charsets that don't come up directly in email but are important for implementing the entire suite of protocols: UTF-7, used in IMAP (more on that later), and Punycode (more on that later, too).
The suite of character sets falls into three main categories. First is the set of fixed-width character sets, most notably ASCII and the ISO 8859 suite of charsets, as well as UCS-2 (2 bytes per character) and UTF-32 (4 bytes per character). Since the major East Asian languages are all ideographic, which require a rather large number of characters to be encoded, fixed-width character sets are infeasible. Instead, many choose to do a variable-width encoding: Shift-JIS lets some characters (notably ASCII characters and half-width katakana) remain a single byte and uses two bytes to encode all of its other characters. UTF-8 can use between 1 byte (for ASCII characters) and 4 bytes (for non-BMP characters) for a single character. The final set of character sets, such as the ISO 2022 ones, use escape sequences to change the interpretation of subsequent characters. As a result, taking the substring of an encoding string can change its interpretation while remaining valid. This will be important later.
Two more problems related to character sets are worth mentioning. The first is the byte-order mark, or BOM, which is used to distinguish whether UTF-16 is written on a little-endian or big-endian machine. It is also sometimes used in UTF-8 to indicate that the text is UTF-8 versus some unknown legacy encoding. It is also not supposed to appear in email, but I have done some experiments which suggest that people use software that adds it without realizing that this is happening. The second issue, unsurprisingly [7], is that for some character sets (Big5 in particular, I believe), not everyone agrees on how to interpret some of the characters.
The largest problem of internationalization that applies in a general sense is the problem of case insensitivity. The 26 basic Latin letters all map nicely to case, having a single uppercase and a single lowercase variant for each letter. This practice doesn't hold in general—languages like Japanese lack even the notion of case, although it does have two kana variants that hold semantic differences. Rather, there are three basic issues with case insensitivity which showcase enough of its problems to make you want to run away from it altogether [8].
The simplest issue is the Greek sigma. Greek has two lowercase variants of the sigma character: σ and ς (the "final sigma"), but a single uppercase variant, Σ. Thus mapping a string s to uppercase and back to lowercase is not equivalent to mapping s directly to lower-case in some cases. Related to this issue is the story of German ß character. This character evolved as a ligature of a long and short 's', and its uppercase form is generally held to be SS. The existence of a capital form is in some dispute, and Unicode only recently added it (ẞ, if your software supports it). As a result, merely interconverting between uppercase and lowercase versions of a string does not necessarily lead to a simple fixed point. The third issue is the Turkish dotless i (ı), which is the lowercase variant of the ASCII uppercase I character to those who speak Turkish. So it turns out that case insensitivity isn't quite the same across all locales.
Again unsurprisingly in light of the issues, the general tendency towards case-folding or case-insensitive matching in internationalized-aware specifications is to ignore the issues entirely. For example, asking for clarity on the process of case-insensitive matching for IMAP folder names, the response I got was "don't do it." HTML and CSS moved to the cumbersomely-named variant known as "ASCII-subset case-insensitivity", where only the 26 basic Latin letters are mapped to their (English) variants in case. The solution for email is also a verbose variant of "unspecified," but that is only tradition for email (more on this later).
Now that you have a good idea of the general issues, it is time to delve into how the developers of email rose to the challenge of handling internationalization. It turns out that the developers of email have managed to craft one of the most perfect and exquisite examples I have seen of how to completely and utterly fail. The challenges of internationalized emails are so difficult that buggier implementations are probably more common than fully correct implementations, and any attempt to ignore the issue is completely and totally impossible. In fact, the faults of RFC 2047 are my personal least favorite part of email, and implementing it made me change the design of JSMime more than any other feature. It is probably the single hardest thing to implement correctly in an email client, and it is so broken that another specification was needed to be able to apply internationalization more widely (RFC 2231).
The basic problem RFC 2047 sets out to solve is how to reliably send non-ASCII characters across a medium where only 7-bit characters can be reliably sent. The solution that was set out in the original version, RFC 1342, is to encode specific strings in an "encoded-word" format: =?charset?encoding?encoded text?=. The encoding can either be a 'B' (for Base64) or a 'Q' (for quoted-printable). Except the quoted-printable encoding in this format isn't quite the same quoted-printable encoding used in bodies: the space character is encoded via a '_' character instead, as spaces aren't allowed in encoded-words. Naturally, the use of spaces in encoded-words is common enough to get at least one or two bugs filed a year about Thunderbird not supporting it, and I wonder if this subtle difference between two quoted-printable variants is what causes the prevalence of such emails.
One of my great hates with regard to email is the strict header line length limit. Since the encoded-word form can get naturally verbose, particularly when you consider languages like Chinese that are going to have little whitespace amenable for breaking lines, the ingenious solution is to have adjacent encoded-word tokens separated only by whitespace be treated as the same word. As RFC 6857 kindly summarizes, "whitespace behavior is somewhat unpredictable, in practice, when multiple encoded words are used." RFC 6857 also suggests that the requirement to limit encoded words to only 74 characters in length is also rather meaningless in practice.
A more serious problem arises when you consider the necessity of treating adjacent encoded-word tokens as a single unit. This one is so serious that it reaches the point where all of your options would break somebody. When implementing an RFC 2047 encoding algorithm, how do you write the code to break up a long span of text into multiple encoded words without ever violating the specification? The naive way of doing so is to encode the text once in one long string, and then break it into checks which are then converted into the encoded-word form as necessary. This is, of course, wrong, as it breaks two strictures of RFC 2047. The first is that you cannot split the middle of multibyte characters. The second is that mode-switching character sets must return to ASCII by the end of a single encoded-word [9]. The smarter way of building encoded-words is to encode words by trying to figure out how much text can be encoded before needing to switch, and breaking the encoded-words when length quotas are exceeded. This is also wrong, since you could end up violating the return-to-ASCII rule if your don't double-check your converters. Also, if UTF-16 is used as the basis for the string before charset conversion, the encoder stands a good chance of splitting up creating unpaired surrogates and a giant mess as a result.
For JSMime, the algorithm I chose to implement is specific to UTF-8, because I can use a property of the UTF-8 implementation to make encoding fast (every octet is looked at exactly three times: once to convert to UTF-8, once to count to know when to break, and once to encode into base64 or quoted-printable). The property of UTF-8 is that the second, third, and fourth octets of a multibyte character all start with the same two bits, and those bits never start the first octet of a character. Essentially, I convert the entire string to a binary buffer using UTF-8. I then pass through the buffer, keeping counters of the length that the buffer would be in base64 form and in quoted-printable form. When both counters are exceeded, I back up to the beginning of the character, and encode that entire buffer in a word and then move on. I made sure to test that I don't break surrogate characters by making liberal use of the non-BMP character U+1F4A9 [10] in my encoding tests.
The sheer ease of writing a broken encoder for RFC 2047 means that broken encodings exist in the wild, so an RFC 2047 decoder needs to support some level of broken RFC 2047 encoding. Unfortunately, to "fix" different kinds of broken encodings requires different support for decoders. Treating adjacent encoded-words as part of the same buffer when decoding makes split multibyte characters work properly but breaks non-return-to-ASCII issues; if they are decoded separately the reverse is true. Recovering issues with isolated surrogates is at best time-consuming and difficult and at worst impossible.
Yet another problem with the way encoded-words are defined is that they are defined as specific tokens in the grammar of structured address fields. This means that you can't hide RFC 2047 encoding or decoding as a final processing step when reading or writing messages. Instead you have to do it during or after parsing (or during or before emission). So the parser as a result becomes fully intertwined with support for encoded-words. Converting a fully UTF-8 message into a 7-bit form is thus a non-trivial operation: there is a specification solely designed to discuss how to do such downgrading, RFC 6857. It requires deducing what structure a header has, parsing that harder, and then reencoding the parsed header. This sort of complicated structure makes it much harder to write general-purpose email libraries: the process of emitting a message basically requires doing a generic UTF-8-to-7-bit conversion. Thus, what is supposed to be a more implementation detail of how to send out a message ends up permeating the entire stack.
Unfortunately, the developers of RFC 2047 were a bit too clever for their own good. The specification limits the encoded-words to occurring only inside of phrases (basically, display names for addresses), unstructured text (like the subject), or comments (…). I presume this was done to avoid requiring parsers to handle internationalization in email addresses themselves or possibly even things like MIME boundary delimiters. However, this list leaves out one common source of internationalized text: filenames of attachments. This was ultimately patched by RFC 2231.
RFC 2231 is by no means a simple specification, since it attempts to solve three problems simultaneously. The first is the use of non-ASCII characters in parameter values. Like RFC 2047, the excessively low header line length limit causes the second problem, the need to wrap parameter values across multiple line lengths. As a result, the encoding is complicated (it takes more lines of code to parse RFC 2231's new features alone than it does to parse the basic format [11]), but it's not particularly difficult.
The third problem RFC 2231 attempts to solve is a rather different issue altogether: it tries to conclusively assign a language tag to the encoded text and also provides a "fix" for this to RFC 2047's encoded-words. The stated rationale is to be able to have screen readers read the text aloud properly, but the other (much more tangible) benefit is to ameliorate the issues of Unicode's Han unification by clearly identifying if the text is Chinese, Japanese, or Korean. While it sounds like a nice idea, it suffers from a major flaw: there is no way to use this data without converting internal data structures from using flat strings to richer representations. Another issue is that actually setting this value correctly (especially if your goal is supporting screen readers' pronunciations) is difficult if not impossible. Fortunately, this is an entirely optional feature; though I do see very little email that needs to be concerned about internationalization, I have yet to find an example of someone using this in the wild.
If you're the sort of person who finds properly writing internationalized text via RFC 2231 or RFC 2047 too hard (or you don't realize that you need to actually worry about this sort of stuff), and you don't want to use any of the several dozen MIME libraries to do the hard stuff for you, then you will become the bane of everyone who writes email clients, because you've just handed us email messages that have 8-bit text in the headers. At which point everything goes mad, because we have no clue what charset you just used. Well, RFC 6532 says that headers are supposed to be UTF-8, but with the specification being only 19 months old and part of a system which is still (to my knowledge) not supported by any major clients, this should be taken with a grain of salt. UTF-8 has the very nice property that text that is valid UTF-8 is highly unlikely to be any other charset, even if you start considering the various East Asian multibyte charsets. Thus you can try decoding under the assumption that is UTF-8 and switch to a designated fallback charset if decoding fails. Of course, knowing which designated fallback to use is a different matter entirely.
Stepping outside email messages themselves, internationalization is still a concern. IMAP folder names are another well-known example. RFC 3501 specified that mailbox names should be in a modified version of UTF-7 in an awkward compromise. To my knowledge, this is the only remaining significant use of UTF-7, as many web browsers disabled support due to its use in security attacks. RFC 6855, another recent specification (6 months old as of this writing), finally allows UTF-8 mailbox names here, although it too is not yet in widespread usage.
You will note missing from the list so far is email addresses. The topic of email addresses is itself worthy of lengthy discussion, but for the purposes of a discussion on internationalization, all you need to know is that, according to RFCs 821 and 822 and their cleaned-up successors, everything to the right of the '@' is a domain name and everything to the left is basically an opaque ASCII string [12]. It is here that internationalization really runs headlong into an immovable obstacle, for the email address has become the de facto unique identifier of the web, and everyone has their own funky ideas of what an email address looks like. As a result, the motto of "be liberal in what you accept" really breaks down with email addresses, and the amount of software that needs to change to accept internationalization extends far beyond the small segment interested only in the handling of email itself. Unfortunately, the relative newness of the latest specifications and corresponding lack of implementations means that I am less intimately familiar with this aspect of internationalization. Indeed, the impetus for this entire blogpost was a day-long struggle with trying to ascertain when two email addresses are the same if internationalized email address are involved.
The email address is split nicely by the '@' symbol, and internationalization of the two sides happens at two different times. Domains were internationalized first, by RFC 3490, a specification with the mouthful of a name "Internationalizing Domain Names in Applications" [13], or IDNA2003 for short. I mention the proper name of the specification here to make a point: the underlying protocol is completely unchanged, and all the work is intended to happen at roughly the level of getaddrinfo—the internal DNS resolver is supposed to be involved, but the underlying DNS protocol and tools are expected to remain blissfully unaware of the issues involved. That I mention the year of the specification should tell you that this is going to be a bumpy ride.
An internationalized domain name (IDN for short) is a domain name that has some non-ASCII characters in it. Domain names, according to DNS, are labels terminated by '.' characters, where each label may consist of up to 63 characters. The repertoire of characters are the ASCII alphanumerics and the '-' character, and labels are of course case-insensitive like almost everything else on the Internet. Encoding non-ASCII characters into this small subset while meeting these requirements is difficult for other contemporary schemes: UTF-7 uses Base64, which means 'A' and 'a' are not equivalent; percent-encoding eats up characters extremely quickly. So IDN use a different specification for this purpose, called Punycode, which allows for a dense but utterly unreadable encoding. The basic algorithm of encoding an IDN is to take the input string, apply case-folding, normalize using NFKC, and then encode with Punycode.
Case folding, as I mentioned several paragraphs ago, turns out to have some issues. The ß and ς characters were the ones that caused the most complaints. You see, if you were to register, say, www.weiß.de, you would actually be registering www.weiss.de. As there is no indication of Punycode involved in the name, browsers would show the domain in the ASCII variant. One way of fixing this problem would be to work with browser vendors to institute a "preferred name" specification for websites (much like there exists one for the little icons next to page titles), so that the world could know that the proper capitalization is of course www.GoOgle.com instead of www.google.com. Instead, the German and Greek registrars pushed for a change to IDNA, which they achieved in 2010 with IDNA2008.
IDNA2008 is defined principally in RFCs 5890-5895 and UTS #46. The principal change is that the normalization step no longer exists in the protocol and is instead supposed to be done by applications, in a possibly locale-specific manner, before looking up the domain name. One reason for doing this was to eliminate the hard dependency on a specific, outdated version of Unicode [14]. It also helps fix things like the Turkish dotless I issue, in theory at least. However, this different algorithm causes some domains to be processed differently from IDNA2003. UTS #46 specifies a "compatibility mode" which changes the algorithm to match IDNA2003 better in the important cases (specifically, ß, ς, and ZWJ/ZWNJ), with a note expressing the hope that this will eventually become unnecessary. To handle the lack of normalization in the protocol, registrars are asked to automatically register all classes of equivalent domain names at the same time. I should note that most major browsers (and email clients, if they implement IDN at all) are still using IDNA2003: an easy test of this fact is to attempt to go to ☃.net, which is valid under IDNA2003 but not IDNA2008.
Unicode text processing is often vulnerable to an attack known as the "homograph attack." In most fonts, the Greek omicron and the Latin miniscule o will be displayed in exactly the same way, so an attacker could pretend to be from, say, Google while instead sending you to Gοogle—I used Latin in the first word and Greek in the second. The standard solution is to only display the Unicode form (and not the Punycode form) where this is not an issue; Firefox and Opera display Unicode only for a whitelist of registrars with acceptable polices, Chrome and Internet Explorer only permits scripts that the user claims to read, and Safari only permits scripts that don't permit the homograph attack (i.e., not Cyrillic or Greek). (Note: this information I've summarized from Chromium's documentation; forward any complaints of out-of-date information to them).
IDN satisfies the needs of internationalizing the second half of an email address, so a working group was commissioned to internationalize the first one. The result is EAI, which was first experimentally specified in RFCs 5335-5337, and the standards themselves are found in RFCs 6530-6533 and 6855-6858. The primary difference between the first, experimental version and the second, to-be-implemented version is the removal of attempts to downgrade emails in the middle of transit. In the experimental version, provisions were made to specify with every internalized address an alternate, fully ASCII address to which a downgraded message could be sent if SMTP servers couldn't support the new specifications. These were removed after the experiment found that such automatic downgrading didn't work as well as hoped.
With automatic downgrading removed from the underlying protocol, the onus is on people who generate the emails—mailing lists and email clients—to figure out who can and who can't receive messages and then downgrade messages as appropriate for the recipients of the message. However, the design of SMTP is such that it is impossible to automatically determine if the client can receive these new kinds of messages. Thus, the options are to send them and hope that it works or to rely on the (usually clueless) user to inform you if it works. Clearly an unpalatable set of options, but it is one that can't be avoided due to protocol design.
The largest change of EAI is that the local parts of addresses are specified as a sequence of UTF-8 characters, omitting only the control characters [15]. The working group responsible for the specification adamantly refused to define a Unicode-to-ASCII conversion process, and thus a mechanism to make downgrading work smoothly, for several reasons. First, they didn't want to specify a prefix which could change the meaning of existing local-parts (the structure of local-parts is much less discoverable than the structure of all domain names). Second, they felt that the lack of support for displaying the Unicode variants of Punycode meant that users would have a much worse experience. Finally, the transition period would be hopefully short (although messy), so designing a protocol that supports that short period would worsen it in the long term. Considering that, at the moment of writing, only one of the major SMTP implementations has even a bug filed to support it, I think the working group underestimates just how long transition periods can take.
As far as changes to the message format go, that change is the only real change, considering how much effort is needed to opt-in. Yes, headers are now supposed to be UTF-8, but, in practice, every production MIME parser needs to handle 8-bit characters in headers anyways. Yes, message/global can have MIME encoding applied to it (unlike message/rfc822), but, in practice, you already need to assume that people are going to MIME-encode message/rfc822 in violation of the specification. So, in practice, the changes needed to a parser are to add message/global as an alias to message/rfc822 [16] and possibly tweaking some charset detection heuristics to prefer UTF-8. I would very much have liked the restriction on header line length removed, but, alas, the working group did not feel moved to make those changes. Still, I look forward to the day when I never have to worry about encoding text into RFC 2047 encoded-words.
IMAP, POP, and SMTP are also all slightly modified to take account of the new specifications. Specifically, internationalized headers are supposed to be opt-in only—SMTP are supposed to reject sending to these messages if it doesn't support them in the first place, and IMAP and POP are supposed to downgrade messages when requested unless the client asks for them to not be. As there are no major server implementations yet, I don't know how well these requirements will be followed, especially given that most of the changes already need to be tolerated by clients in practice. The experimental version of internationalization specified a format which would have wreaked havoc to many current parsers, so I suspect some of the strict requirements may be a holdover from that version.
And thus ends my foray into email internationalization, a collection of bad solutions to hard problems. I have probably done a poor job of covering the complete set of inanities involved, but what I have covered are the ones that annoy me the most. This certainly isn't the last I'll talk about the impossibility of message parsing either, but it should be enough at least to convince you that you really don't want to write your own message parser.
[1] Date/time, numbers, and currency are the other major aspects of internalization.
[2] I am a native English speaker who converses with other people almost completely in English. That said, I can comprehend French, although I am not familiar with the finer points that come with fluency, such as collation concerns.
[3] C and C++ have a built-in internationalization and localization API, derived from POSIX. However, this API is generally unsuited to the full needs of people who actually care about these topics, so it's not really worth mentioning.
[4] The basic algorithm to encode RFC 2047 strings for any charset are to try to shift characters into the output string until you hit the maximum word length. If the internal character set for Unicode conversion is UTF-16 instead of UTF-32 and the code is ignorant of surrogate concerns, then this algorithm could break surrogates apart. This is exactly how the bug is triggered in Thunderbird.
[5] I'm not discussing Han unification, which is arguably the single most controversial aspect of Unicode.
[6] Official list here means the official set curated by IANA as valid for use in the charset="" parameter. The actual set of values likely to be acceptable to a majority of clients is rather different.
[7] If you've read this far and find internationalization inoperability surprising, you are either incredibly ignorant or incurably optimistic.
[8] I'm not discussing collation (sorting) or word-breaking issues as this post is long enough already. Nevertheless, these also help very much in making you want to run away from internationalization.
[9] I actually, when writing this post, went to double-check to see if Thunderbird correctly implements return-to-ASCII in its encoder, which I can only do by running tests, since I myself find its current encoder impenetrable. It turns out that it does, but it also looks like if we switched conversion to ICU (as many bugs suggest), we may break this part of the specification, since I don't see the ICU converters switching to ASCII at the end of conversion.
[10] Chosen as a very adequate description of what I think of RFC 2047. Look it up if you can't guess it from context.
[11] As measured by implementation in JSMime, comments and whitespace included. This is biased by the fact that I created a unified lexer for the header parser, which rather simplifies the implementation of the actual parsers themselves.
[12] This is, of course a gross oversimplification, so don't complain that I'm ignoring domain literals or the like. Email addresses will be covered later.
[13] A point of trivia: the 'I' in IDNA2003 is expanded as "Internationalizing" while the 'I' in IDNA2008 is for "Internationalized."
[14] For the technically-minded: IDNA2003 relied on a hard-coded list of banned codepoints in processing, while IDNA2008 derives its lists directly from Unicode codepoint categories, with a small set of hard-coded exceptions.
[15] Certain ASCII characters may require the local-part to be quoted, of course.
[16] Strictly speaking, message/rfc822 remains all-ASCII, and non-ASCII headers need message/global. Given the track record of message/news, I suspect that this distinction will, in practice, not remain for long.