and only run rbs tests for ruby 4

and only add pstore do explicit depset in ruby 4, when it leaves the standard gems pack

these get frozen after the first TLS connection attempt performed inside
an SSLSocket.

This is circumvented by passing the same proc at init instead, which
then points at an internal ivar which can be changed.

rbs logger is ractor-unfriendly

the next version of http-2 will support a :closing state, which for all
the functionality supported by httpx should behave as :closed.
`#closed?` returns true for both states, so should be the correct wethod
to query.

the expires at may only be calculated after the token has been generated, specially in situations where the token carries that information and expires_at calls a callback to fetch it

init_time could be calculated at any point during a reconnection, which
could happen before it'd be back to the pool, thereby potentially
leaving to open a window for it, and account millisecond requests as
having taken hours.

this may happen as part of error handling; without this, the resolver
may be left in an inconsistent state where it's open but has no io,
which breaks selection across the board.

there seems to be an issue using registry images in CI.

when it's dynamic. this should tell the end user whether it should check
for existing token freshness or force generation of the token, to avoid
cases where the former may suffer from clock skew and has already failed
once despite being considered fresh.

it's always true on the retry of the first (or the single) request of
the batch, false otherwise.

this caused issues during after fork callbacks, where socket raised an error when writing the termination handshake to, leaving the child process in a corrupt state

this is so that they do not retain it when going back to the pool, because as they're checked back out, the requests sent through it may set its own init_time with that timestamp, which is quite far back in time

also, probe-close spans in datadog on reset, just in case they're left behind

in order not to trigger a retry if using the persistent session in tests

without this, it was possible to abuse  calls and just pass anything, which would be silenty ignored

This reverts commit b1b8d9b8.

if the request came right after a ping, and the error was a ping timeout
error, then the retry should be immediate. previously, this error wasn't
marked as a reconnectable error, so it was skipped.

in http/1.1 connections, the ping path is skipped, and instead, the connection is reconnected, so the request goes through a fresh connection.

however, the ping timeout timer was still being set, and could effectively fail a response still waiting for a valid request

response can be anything, including a webmock response

the proxy plugin already has a conditional plugin for retries, so it best lives there

anything else should just fail hard

persistent: allow certain errors (like http2 goaway error) to be retried without affecting request attempts

these errors are a known clean message from the peer to reattempt; this fixes an issue where, if multiple stale connections are in the pool, and a request gets attempted on them, and it'd fail on them with the same error, then it'd fail without ever having a clean attempt

this allows probing all of them, until a fresh one is found or a new one is created, and then it'd manage to get a fair attempt

this is so that they do not retain it when going back to the pool,
because as they're checked back out, the requests sent through it
may set its own init_time with that timestamp, which is quite far
back in time.

also, probe-close spans in datadog on reset, just in case they're left
behind

being a tad bit more accurate

for friendly object shape

this avoids the emission of the response event two times in the error case, and allows the follow request to just return cleanly, as it has a response

the approach of erroring on something people have been accidentally
relying on is too strict in hindsight for a patch version. this was
reverted, and instead, a deprecation warning will be printed, to give
time for folks to migrate.

the format of options messages (exceptions and warnings) also slightly
changes to wrap the option in quotes, for proper highlighting.