Proposed resolution:

This wording is relative to N4878.

1. Modify [time.duration.io] as indicated:

   template<class charT, class traits, class Rep, class Period, class Alloc = allocator<charT>>
     basic_istream<charT, traits>&
       from_stream(basic_istream<charT, traits>& is, const charT* fmt,
                   duration<Rep, Period>& d,
                   basic_string<charT, traits, Alloc>* abbrev = nullptr,
                   minutes* offset = nullptr);
   

   -3- Effects: Attempts to parse the input stream is into the duration d using the format flags given in the NTCTS fmt as specified in [time.parse]. If the parse parses everything specified by the parsing format flags without error, and yet none of the flags impacts a duration, d will be assigned a zero valueIf the parse fails to decode a valid duration, is.setstate(ios_base::failbit) is called and d is not modified. If %Z is used and successfully parsed, that value will be assigned to *abbrev if abbrev is non-null. If %z (or a modified variant) is used and successfully parsed, that value will be assigned to *offset if offset is non-null.

[ 2021-06-07 Approved at June 2021 virtual plenary. Status changed: Voting → WP. ]

[ 2021-04-20; Reflector poll ]

Set status to Tentatively Ready after eight votes in favour during reflector poll.

The duration specialization of from_stream says in N4878 [time.duration.io]/3:

If the parse parses everything specified by the parsing format flags without error, and yet none of the flags impacts a duration, d will be assigned a zero value.

This is in contrast to the other specializations that say, for example, [time.cal.day.nonmembers]/8:

If the parse fails to decode a valid day, is.setstate(ios_base::failbit) is called and d is not modified.

The wording ("none of the flags impacts a duration" vs. "parse fails to decode a valid [meow]") and semantics ("assigned a zero value" vs. "not modified") are different, and it's not clear why that should be so. It also leaves unspecified what should be done in case of a parse failure, for example parsing "%j" from a stream containing "meow". [time.parse]/12 says that failbit should be set, but neither it nor [time.duration.io]/3 mention the duration result if parsing fails.

This has been discussed at the Microsoft STL project, where Howard Hinnant, coauthor of P0355R7 that added these functions, commented:

This looks like a bug in the standard to me, due to two errors on my part.

I believe that the standard should clearly say that if failbit is set, the parsed variable (duration, time_point, whatever) is not modified. I mistakenly believed that the definition of unformatted input function covered this behavior. But after review, I don't believe it does. Instead each extraction operator seems to say this separately.

I also at first did not have my example implementation coded to leave the duration unchanged. So that's how the wording got in in the first place. Here's the commit where I fixed my implementation: HowardHinnant/date@d53db7a. And I failed to propagate that fix into the proposal/standard.

It would be clearer and simpler for users if the from_stream specializations were consistent in wording and behavior.

Thanks to Stephan T. Lavavej, Miya Natsuhara, and Howard Hinnant for valuable investigation and discussion of this issue.