fix: Close the connection before reporting a stream error, and keep the reconnect loop alive

[jsonbailey]

BEGIN_COMMIT_OVERRIDE
fix: Close the connection before passing a stream error to the error handler
fix: Pass at most one error to the error handler for each connection attempt
fix: Keep reconnecting when closing the previous connection fails
fix: Prevent a reconnect delay of NaN when the retry interval is zero
END_COMMIT_OVERRIDE

Summary

Fixes found while reviewing the Ruby server SDK's RETRY work (ruby-server-sdk#448). That SDK waits inside on_error before it reconnects, which exposed these paths in the reconnect loop:

• Close before reporting. run_stream called on_error before reset_http. A handler that blocks kept the dead connection open for the whole wait, for example after a read timeout. The connection is now closed first.
• One error per attempt. A failure in reset_http was sent to on_error as a second error for the same attempt. It is now logged at warn.
• The worker cannot stop silently. close_connection at the top of the loop was outside any rescue, so a raise ended the worker thread and the client never reconnected. It is now logged and the loop continues.
• No NaN in the backoff. With a Float 0.0 base (for example after a server retry: 0), 0.0 * 2**n became NaN after about 1024 attempts, and [NaN, max].min raised on every later attempt. A zero base now returns no delay, and the attempt count stops growing at the maximum.
• Two comments named StreamingHTTPConnection, which was removed in 2.0.0. They now name HTTP::Response.

No API change.

Testing

• New specs for each change. Each fails when its fix is reverted.
• bundle exec rspec spec: 182 examples, 0 failures. bundle exec rubocop: no offenses.

---

Note

Overview
Hardens the SSE client reconnect loop so cleanup and error reporting happen in a safer order, and fixes backoff edge cases that could break long-running reconnects.

In run_stream, the connection is now closed (reset_http) before on_error runs, so blocking error handlers do not keep a dead socket open. Close failures are handled via log_close_errors: they are logged only (not sent to on_error), and close_connection at the start of each loop is wrapped the same way so a close raise no longer kills the worker thread. Backoff#next_interval returns 0 when the base or max interval is non-positive (avoids NaN after many attempts with a 0.0 base from retry: 0), and stops incrementing the attempt counter once the delay hits the max ceiling.

New specs cover close-before-error ordering, a single error per attempt when close fails, worker survival, and backoff behavior. Comments now refer to HTTP::Response instead of the removed StreamingHTTPConnection. No public API changes.

^{Reviewed by Cursor Bugbot for commit dedf9f8. Bugbot is set up for automated code reviews on this repo. Configure here.}