Fixes #1478

Problem

onClosed was implemented as:

this._writer.closed.then(cb).catch(_noop);

When the client disconnects, the readable side of the TransformStream is cancelled, which puts the writable side into an errored state — so writer.closed rejects and the callback is routed into .catch(_noop), never running. It only fired on a graceful close().

Autoclose was additionally wired exclusively to the Node runtime (runtime?.node?.res?.once("close", ...)), so on web-standard runtimes (Deno, Bun, Workerd, service workers, plain app.fetch() consumers) a client disconnect neither closed the stream nor ran cleanup. Timers, DB handles, and generator loops leaked per connection.

Fix

• Close callbacks are queued and fired from the single _writer.closed.catch(_noop).finally(...) handler, so they run for both graceful close and cancellation-induced rejection.
• That handler marks the stream disposed when autoclose !== false, so a disconnect stops further pushes on every runtime, not just Node.
• onClosed registered after the stream already closed now fires on a microtask instead of silently never running.
• Callback throws stay swallowed, matching the previous .catch(_noop) behavior.
• The Node res listener is kept as a belt-and-braces signal for the Node path.

Also in this PR

Two adjacent defects in the same teardown path, each with its own commit:

Async onClosed callbacks leaked rejections. Swallowing only synchronous throws left an async callback that rejects to surface as an unhandled rejection, and it aborted the remaining callbacks in the queue. Callbacks are now invoked through a helper that absorbs both, so one bad callback can no longer take down the process or starve the ones behind it.

close() dropped data buffered while paused. close() never flushed _unsentData, and flush() short-circuits once the stream is closed — so anything queued by a push() during pause() was unrecoverable. A handler that paused for backpressure, queued a final message and then closed left the client with a cleanly ended stream and a silently missing message. close() now unpauses and flushes before closing the writer. This adds no new blocking, since _writer.close() already awaited the queue draining.

Test

test/sse.test.ts — the client reads one chunk, cancels the reader, and asserts onClosed fires. Under describeMatrix it failed on web and passed on node before the fix (exactly the reported asymmetry); both pass now.

The two follow-up fixes are covered by their own regression tests — an integration test per matrix target plus a unit test for the paused-close flush, all failing before their respective commits. Full suite green.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Improved server-sent event (SSE) stream cleanup and lifecycle handling on client disconnects.
  • onClosed handlers now fire consistently, including after disconnects and when closing a paused stream.
  • Callbacks registered after the stream is already closed run promptly, even if earlier handlers throw or reject.
  • Closing a paused SSE stream now flushes buffered output so no messages are missed.
• Tests
  • Added/updated SSE and unit tests covering disconnect behavior, paused-close flushing, and onClosed error scenarios.