v13.2.0

Added

• HostedRelayWakeResult has an optional retainedWork field. It copies the
  Relay report of the wake pull. True means that the Relay keeps a message
  after the page, or a Relay receipt that the page did not carry, so run
  another pullHostedRelayAfterWake() pass. False means that the page carried
  all retained work. A Relay that does not report it gives no field. Then the
  retained work is unknown, so do not read an absent field as false.

Fixed

• A hosted mailbox subscription now acknowledges each recovery pull page in
  one HTTP request, also while the socket is live, and waits for the response.
  When the page was full, with 100 messages or 32 Relay receipts, and the
  Relay removed work or issued a cursor, the next page follows at once. It
  does not wait for a timer. So a backlog of Relay receipts drains on the same
  socket, with no reconnect. A socket frame never starts a pull.
• A lost acknowledgment request or a lost acknowledgment response now starts
  one pull from the mailbox head after a wait. The wait starts at 2 seconds
  and doubles after each further loss, up to 64 seconds. An acknowledgment
  whose requests all succeed starts it at 2 seconds again. A pull that is
  requested while the acknowledgment fails waits for that head pull. So a
  lasting acknowledgment failure backs off. A replayed Relay receipt is not
  handed off again.
• A subscription now keeps at most 132 Relay receipts in memory while a
  receipt listener is slow. One batch takes at most 32 receipts. When the
  queue of 100 items is full, the rest stay in the Relay, and the subscription
  pulls them when the queue is empty, with no wait.
• A pull page now decodes the optional failedEntries list of the Relay. It
  holds each retained message that the Relay read and could not serve. The
  subscription never acknowledges such a message. It pulls again from the
  mailbox head after the failed-work wait, which starts at 2 seconds and
  doubles up to 64 seconds. A page pulled with a cursor does not start that
  wait again, so a long backlog does not cut the wait short. A page with a
  failedEntries value that is not a list of entries is refused. A page that
  holds only failed entries keeps its nextCursor, so the subscription and
  pullHostedRelayAfterWake() continue past it with the cursor. The
  subscription pulls such cursor pages to their end before the pull from the
  head. It does not continue on a nextCursor that equals the cursor of the
  pull.
• A recovery pull page whose retainedWork is true now continues at once
  when the Relay removed work or issued a cursor, also when the page is not
  full. A page without retainedWork uses the full-page rule.
• A subscription now keeps at most one HTTP acknowledgment in flight, and a
  pull waits for its response. When a socket burst ends and a pull follows,
  the Relay receipts of the burst are acknowledged over HTTP before the pull.
  So the pull page does not offer those receipts again, and the next pages
  follow.
• A pull that a subscription requests while a pull fails now waits for the
  retry schedule of the failed pull. It does not run at once after the
  failure.