0403 — Compaction's winner rule generalized to a signed caller-supplied rank compared before id

Context. Compaction’s rule was “arrival order wins.” A bridge consumer re-producing old data into a live topic races that topic’s own producers: whichever write lands last wins, so a stale backfill landing after a fresh live write would silently overwrite it.

Decision. CompactionRank BIGINT NOT NULL DEFAULT 0 on message_log and compaction_head (the winner’s rank is stored, not just its id), set via ProduceOptions.CompactionRank. Both upsert guards — pkg/producer/datastore.go’s protectedInsertSQL and batch.go’s batch path — became the native row compare (compaction_head.compaction_rank, compaction_head.head_id) < (EXCLUDED..., EXCLUDED...). The rank is signed so the bridge can write below the default: bridge writes at rank −1, live producers at their default rank 0, and a live write always wins regardless of which one the database committed first.

Consequences. The migration race becomes a declarative comparison instead of a timing outcome. All-default traffic reproduces the old semantics bit-for-bit, and every losing row stays physically present, just never claimed. examples/phase_1/compactionranklab proves the pin, the two-arrival-order bridge race, and the losing-row retention.