@jwakely , the suggestion seems to be eminently reasonable. What do you think?

is a general rule about many operations on a specific container more specific than a more descriptive rule about a specific member function of many containers, particularly one which does specify an exception?

It's as you just pointed out with std::string. Explicitly describing a member function is not license to deviate from the container requirements.

The container requirements centralize exceptions. Elsewhere, where some [library] requirement is overriden, the exception is at the point of use, like

4
#
Note B: Notwithstanding the provisions of [meta.type.synop], and pursuant to [namespace.std], a program may specialize common_­type<T1, T2> for types T1 and T2 such that is_­same_­v<T1, decay_­t> and is_­same_­v<T2, decay_­t> are each true.

Hmm, N3108 solved this problem, but then that paragraph was removed in WP N3242 (see the red change marks in [basic.string] p3). I can't find any justification for removing that text. The editor's report in N3243 doesn't mention it. I can't find any paper or LWG issue saying to remove it.

We had to put back the wording about basic_string not using the allocator's construct and destroy members in P1072R10. That should have been unnecessary, because we already had that wording prior to N3242.

LWG1415 described an overlapping issue, and was resolved by N3108, which added normative wording exempting basic_string from the allocator-aware construct and destroy requirements as well as a footnote stating this exception for swap, but notes are non-normative.

That's not quite right. N3108 did not add the text about construct and destroy, it added the text shown in bold in N3108, which includes a normative exemption for swap (and a footnote saying why the exemption exists).

N3108 solved this problem, then some mystery edit in N3242 removed it (and the allocator stuff) and somehow nobody noticed. I'm quite concerned by this (LWG 2263 suggests there was another accidental deletion in N3242).

The simplest resolution is to simply add 'and basic_string' after 'other than array' in [container.reqmts] p65.

I don't think that's sufficient. That would allow us to copy characters when swapping strings, but it's not clear to me that "other than array" in that first sentence also exempts array from the penultimate sentence in the paragraph, about iterators. And if exceptions in that first sentence to apply to the whole paragraph, then your suggestion would mean that basic_string no longer uses propogate_on_container_swap, which would be wrong. We also have the following in [container.reqmts] p16 (16.6) which should exclude array:

no swap() function invalidates any references, pointers, or iterators referring to the elements of the
containers being swapped.

Technically that only applies to containers in "this Clause", so already excludes basic_string, but that's subtle and it's not clear to me why those requirements don't apply to basic_string.

I think we need an LWG issue, this is messy.

That old Working Draft was just before we moved to git, so we can't really trace what caused the removal.