release: 0.16.0 (#325)

* docs(stubs): carry the current API into the IDE stubs

The hand-kept mirror lacked Room, RoomDeliveryException and 14 methods
(HttpServer rooms, the reliable-send retry settings, tryWriteMessage(),
trySseEvent()), and 15 docblocks had fallen behind the stubs since the
last full pass at 0.11.0: setHeader(), setStatusCode(), write(),
getRuntimeStats() among them. Signatures now match the stubs one for
one. readBody() names the HttpException a refused body throws.

* release: 0.16.0

release: 0.15.0 (#316)

* docs(stubs): carry this release's API into the IDE stubs

The mirror is hand-maintained and lagged behind two changes shipping in
0.15.0: hide() documented two of the five pattern shapes it reads, and
getBoundListeners() was absent along with what port 0 means for
addListener().

* release: 0.15.0

Two fixes, both visible to a handler, and both about an answer the se…

…rver gave

that was not the one it meant. No breaking change; a build on 0.13.0 upgrades
without touching a line.

- **`awaitWritable()` waits for the stream window on HTTP/3 (#265).**
  `h3_stream_ops` declared no `wait_writable`, and `awaitWritable()` answers
  false for a transport that cannot wait — the answer that keeps a handler from
  being told "go ahead" and spinning without yielding. It was the wrong answer
  here: the HTTP/3 stream already owned the wake path, and HTTP/1, HTTP/2 and
  the reactor-pool worker all park on theirs. A handler was left with a
  sleep-and-retry loop or a blocking `write()`, and on a pool worker that loop
  holds the thread carrying the other requests. The wait is bounded by the
  shorter of the caller's timeout and the connection's write deadline, the rule
  HTTP/2 already followed.

- **An aborted response is labelled by the call that finished it (#267).** One
  field answered two questions. `abort()` on a stream the transport had put no
  byte of — `sseStart()` with no event is the ordinary way there — left the
  response in the state a clean `end()` leaves, because the label was set only
  where the transport reported a reset. So the second `abort()`, documented as a
  no-op, threw "the body was already finished cleanly", and a later `write()`
  named `end()` instead of `abort()`. The natural call site for `abort()` is a
  `catch` or a `finally`, where a throw replaces the diagnosis the handler is
  carrying. The label now records which call finished the response; a separate
  flag records what the peer saw, so the access record and
  `responses_aborted_total` still count only a body that stopped short.

354 of 374 phpt (20 skipped, 0 failed) and 18 of 18 unit cases, against php-src
`true-async` with php-async `main`, PHP 8.6.0-dev ZTS, TrueAsync ABI 0.26.0.
Each fix lands with the test that fails without it: `h3/069` reads `room=0`
against 0.13.0, and `core/063` gains the third state a started stream can be in
and disagrees with 0.13.0 on all three of its new lines.

Merge pull request #264 from true-async/release-0.13.0

release: 0.13.0

docs(changelog): release 0.11.3

docs(changelog): release 0.11.2

docs(changelog): release 0.11.1

docs(changelog): release 0.11.0