fix: correct millis-to-micros conversion for Avro timestamp-millis in Document conversion - #122

Follow-up to #119, in the Avro → Firestore/Datastore Document path.

Problem

DateTimeUtil.toProtoTimestamp(Long) interprets its argument as epoch microseconds:

public static com.google.protobuf.Timestamp toProtoTimestamp(final Long epocMicros) {
    long second = epocMicros / 1000_000L;
    long nano   = epocMicros % 1000_000L * 1000L;

AvroToDocumentConverter passes it epoch seconds on the timestamp-millis branch:

if(LogicalTypes.timestampMillis().equals(schema.getLogicalType())) {
    yield ...setTimestampValue(DateTimeUtil.toProtoTimestamp(l/1000L)).build();   // l is epoch millis
} else if(LogicalTypes.timestampMicros().equals(schema.getLogicalType())) {
    yield ...setTimestampValue(DateTimeUtil.toProtoTimestamp(l)).build();         // correct
}

The value ends up wrong by a factor of 10^6. 2026-08-07T10:15:30.123Z (epoch millis 1786097730123) becomes 1786 seconds — 1970-01-01T00:29:46Z. Every timestamp-millis field written through this path collapses to just after the epoch, silently.

The adjacent timestamp-micros branch passing l straight through confirms microseconds is the expected unit. Of the call sites of toProtoTimestamp outside DateTimeUtil, this is the only one that divides.

The path is reachable from MElement and ElementToDocumentConverter, and timestamp-millis is a first-class logical type here (REQUIRED_LOGICAL_TIMESTAMP_MILLI_TYPE / NULLABLE_LOGICAL_TIMESTAMP_MILLI_TYPE).

Fix

l/1000L → l * 1000L, converting millis to the micros the callee expects.

Tests

Adds AvroToDocumentConverterTest with three cases:

testTimestampMicrosField passes both before and after by design — it guards the already-correct branch against over-correction.

before fix:  Tests run: 3, Failures: 2
after fix:   Tests run: 3, Failures: 0

Note

Building from a clean clone needed org.apache.kafka:kafka-clients added locally — Beam ships it at provided scope so it is not pulled transitively by beam-sdks-java-io-kafka (same thing reported in the second half of #108). That is unrelated to this change and is not included here; I only added it locally to run the tests. Happy to open a separate PR if useful.