:noindex: also drops the signature anchor and permalink, not just the index entry

[ggiesen]

HTTPResource.add_target_and_index() does two things and treats them differently on purpose:

def add_target_and_index(self, name_cls, sig, signode):
    signode['ids'].append(http_resource_anchor(*name_cls[1:]))   # always
    if 'noindex' not in self.options:                            # gated
        self.env.domaindata['http'][self.method][sig] = (...)

i.e. the per-page anchor is unconditional and only the global route registration is gated behind :noindex:. However, Sphinx's ObjectDescription.run() skips the call to add_target_and_index() entirely when the option is set (this outer gate exists at least as far back as Sphinx 2.4.5), so the inner distinction never runs: a :noindex:'d directive loses its #<method>--<path> anchor and the HTML permalink along with the registration, and external deep links into that page land at the top of the page.

Anchors are per-document HTML ids and cannot collide across pages, so keeping them under :noindex: is always safe; only the global registration can legitimately be suppressed.

Motivating case: a project documenting the same routes on more than one page (in our case saltstack/salt, where three netapi modules document overlapping routes). Duplicate registrations are only detected in HTTPDomain.merge_domaindata(), i.e. during parallel reads, so sphinx-build -W -j auto fails nondeterministically depending on where the worker chunk boundary lands between the colliding documents (saltstack/salt#69724). :noindex: on the non-canonical copies is the natural remedy, but today it costs the anchors and permalinks on those pages.

Proposed fix: handle the option in HTTPResource.run() - pop it so the base class still calls add_target_and_index(), keep the anchor, and gate only the registration. PR to follow.