test(replication): verify per-record expiresAt eviction across nodes

[kriszyp]

Summary

• Adds v5→v5 multi-record batch TTL eviction test to replicationTopology suite
• Adds v4→v5 cross-version TTL eviction test to the existing legacy replication test

Background

Field report: records arriving via cross-version replication from v4 peers don't evict on v5 receivers, even past their expiresAtTimestamp. Local writes on the same v5 node evict correctly.

Root causes found during investigation:

1. Harper core (v5) — batched transaction context bug (HarperFast/harper#639, #640)

In a multi-record replication batch, every event after the first is processed with the first event as context. The options object passed to _writeUpdate didn't include expiresAt, and _writeUpdate only read from context, so subsequent records in a batch received no expiration. scheduleCleanup() was also only armed when context.expiresAt was truthy, missing replicated writes.

2. harperdb v4 — table copy encoding bug (harperdb/harperdb#3120)

When a new node joins a v4 cluster, the table copy path calls createAuditEntry with entry.metadataFlags & ~0xff as extendedType. This masks HAS_EXPIRATION (0x10, lower byte) without promoting it to HAS_EXPIRATION_EXTENDED_TYPE (0x1000), so the expiresAt float64 is never encoded into the binary payload. Receivers see auditRecord.expiresAt === undefined for every table-copied record.

Test design

v5→v5 test: Creates a 1-second TTL table on node 0, upserts 4 records in one batch (exercising the batched-txn path), waits for replication confirmation, then asserts all 4 records are gone from every node after TTL+buffer.

v4→v5 test: Pre-creates the table on the v5 cluster with the same expiration value before connecting to the v4 node (workaround for the v4 table-copy encoding bug — the fallback expirationMs + Date.now() on the v5 receiver computes a fresh expiresAt at receipt time). Writes records on the v4 node, waits for replication, asserts eviction.

Dependencies

• Requires HarperFast/harper#639 for the v5→v5 and v4→v5 tests to pass reliably (core submodule bump needed after merge)

---

Signed-off by Claude

[claude]

Reviewed; no blockers found.