Play Framework

fix(server): bound SSE/chunked write queue with backpressure (PF-134)

LazyChunkedInput.nextChunks was an unbounded ConcurrentLinkedQueue: a
producer outpacing a slow client (canonical SSE failure) grew it without
limit, with no channel.isWritable()/watermark check anywhere in the
writeChunk path.

- Bound the queue by queued bytes (8 MiB high / 4 MiB low watermark with
  hysteresis). When over HIGH, the writing virtual thread parks until the
  IO thread drains below LOW; cheap on VTs under JEP 491. Parking is gated
  on !inEventLoop() so a raw-plugin write from channelRead can never block
  the event loop that drains the queue (would deadlock). On close while
  parked, the writer throws so SseStream's close-on-failure path fires.
- readChunk now wraps the byte[] zero-copy (Unpooled.wrappedBuffer) instead
  of allocating + copying a fresh buffer per chunk; the array is never
  mutated after enqueue.
- closed is now volatile (written by invoker VT, read by IO thread).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
(cherry picked from commit 80853cedb687d05c9d16f27fe886ec0eaccb5b6b)