[CCCLL]

Fixes #1531.

ZonedDateTimeSerializer wrote the date, time and zone of a ZonedDateTime and read it with ZonedDateTime.of, which takes the earlier offset in an overlap, when clocks are set back. A ZonedDateTime with the later offset, eg 2026-11-01T01:30-05:00[America/New_York], was read as 2026-11-01T01:30-04:00[America/New_York], an instant one hour earlier, without an exception.

Like for locales with a script or extensions (#1334, #1525), a null zone ID, which Kryo 5 never wrote, now marks an offset. It is written before the zone ID only for a ZonedDateTime with the later offset in an overlap. The offset is read with ZonedDateTime.ofLocal, which uses it if it is valid for the date and time in the zone, and else the earlier offset as before, eg if the zone's rules changed. Other ZonedDateTimes are written as before, so Kryo 6 reads all ZonedDateTimes written by Kryo 5 and the golden files are unchanged. With Format.KRYO5, they are written like Kryo 5 wrote them, without the offset. MIGRATION.md has an entry.

To find out if a ZonedDateTime has the later offset, writing it now looks up the zone's transition for its date and time. In a rough loop on JDK 25, writing a ZonedDateTime with a region zone took about 31 ns instead of 21 ns. Reading is unchanged. Always writing the offset would avoid the lookup, but it would make every ZonedDateTime longer and still need a marker to read the golden files.

Tests: TimeSerializersTest.testZonedDateTimeOverlap round trips both offsets of an overlap and checks that an offset that is not valid for the date and time is not used. Kryo5RoundTripTest.testZonedDateTimeLaterOffsetAtOverlap checks that the Kryo 5 format writes the same bytes as Kryo 5 and that Kryo 5 reads them. The tests fail without the fix, and also when the format check, the read of the offset or the check for the later offset is removed.

I also checked all zones, around every transition from 1900 to 2100 and at random instants, 838,235 values with Output and ByteBufferOutput, 117,381 of them with the later offset in an overlap: Kryo 6 round trips each value, writes the same bytes as Kryo 5 except for the later offsets, and writes the same bytes as Kryo 5 with the Kryo 5 format.

mvn -pl main test passes on JDK 17, 21 and 25 (built with JDK 25), including the Android API check. mvn -pl main formatter:validate passes.

master has the same bug, but the fix changes the format, so this is for kryo-6 only.

[Copilot]

🟢 Approval recommended

No unresolved blocking issues were identified.

No open issues 🎉

What changed in this PR

Preserves the later offset when serializing ZonedDateTime values in overlaps while maintaining Kryo 5 compatibility.

Changes:

• Adds overlap-offset preservation for Kryo 6.
• Adds round-trip and compatibility tests.
• Documents the format change.

File	Description
test/​com/​esotericsoftware/​kryo/​serializers/​TimeSerializersTest.java	Tests overlap round trips and invalid offsets.
test/​com/​esotericsoftware/​kryo/​Kryo5RoundTripTest.java	Verifies Kryo 5 format compatibility.
src/​com/​esotericsoftware/​kryo/​serializers/​TimeSerializers.java	Preserves later overlap offsets.
MIGRATION.md	Documents the format and behavior change.

_{Review effort: Lite}

[theigl]

Thanks again @CCCLL!