The prefix-based field extraction fixes the Unicode-line-separator truncation, but I think one SSE whitespace bug remains in the same path.

At event dispatch, SseLineSubscriber still does:

String eventData = this.eventBuilder.toString();
SseEvent sseEvent =
    new SseEvent(currentEventId.get(), currentEventType.get(), eventData.trim());

Per SSE parsing, each data: field removes at most one optional U+0020 after the colon, data lines are joined with \n, and then only the final synthetic newline is removed. A whole-event .trim() also removes significant whitespace from the first/last data line.

The current new test avoids this by putting the second significant space on an interior line:

data: first
data:  second

so " second" survives. A single-line regression exposes the remaining loss:

List<SseEvent> events = parse(List.of("data:  first  ", ""));
assertThat(events.get(0).data()).isEqualTo(" first  ");

I expect the current PR head still returns "first".

Since this PR is already replacing regex/trim-based extraction with the SSE field rule, it seems worth removing the whole-event .trim() as well and dropping only the trailing separator appended by the subscriber (e.g. remove one final \n). Otherwise U+2028/U+2029/U+0085 are fixed, but valid leading/trailing spaces are still silently changed.

You're absolutely right — thank you for catching this. The trim() was carried over from the old implementation, and my existing tests accidentally avoided it by placing the second space on an interior line, exactly as you described.

Fixed along the lines you suggested: dispatching now only removes the single trailing separator that the data: handler appended (concatenatedDataLines()), so significant leading/trailing whitespace of the first and last data lines is preserved. I also added two regression tests covering your single-line example ("data: padded " → " padded ") and the multi-line first/last whitespace case, and updated the comment on the existing single-leading-space test that implicitly documented the old trim behaviour.

Everything passes locally (7/7 in ResponseSubscribersTest), pushed as a single amended commit.

Verified the updated head. The whole-event .trim() is gone, dispatch now removes only the synthetic final newline, and the two added regressions cover both the single-line and first/last multi-line whitespace cases I was concerned about.

That addresses my review point. Thanks for turning it around quickly.