0618 — the schema version is a message_log column declared by the Message type, not a topic key

Supersedes [0401], [0405], [0408], [0409]; amends [0406].

Context

topic.SchemaVersion(1) appears at every RegisterTopic / Register / GetTopic call and playground scenario 01 flagged it as the parameter nobody can explain on sight. Tracing it back: [0401] solved silent JSON decode drift by making a payload bump a new physical topic, and [0405] then had to key the catalog (name, schema_version) so every verb could say which physical topic it meant. The premise both accepted without naming it is that the row carries no description of its own shape — the library owns the codec (NewProducer[Message] decodes payload into T) yet nothing on the row says what T the bytes are. Bytes-first systems (Kafka, SQS, RabbitMQ) never meet the problem; Confluent’s wire format and Temporal’s Payload.metadata put the schema identity on the message. Postgres peers that own the codec (River) rely on JSON’s additive compatibility and version nothing.

Decision

• message_log.schema_version BIGINT NOT NULL — what this row’s payload is. topic_config.schema_version is deleted; UNIQUE (name).
• The Message type declares its version: Producer[Message topic.Versioned] / Consumer[Message topic.Versioned] where type Versioned interface { SchemaVersion() int } — a plain int, the method name already says what it is, so the old SchemaVersion type is deleted. Every produce writes Message.SchemaVersion() into the row. It is the only source — no config override.
• A consumer group claims only rows whose schema_version equals its Message type’s; the predicate joins the binding predicate in readMessages and the exception claim. Other versions pass under the cursor like a binding miss (skipped, not decoded).
• RegisterTopic / GetTopic / AlterTopic / DestroyTopic / RenameTopic / Producer.Register / Consumer.Register and the CLI --schema-version flag drop the version parameter. Log/metric version attribute now names the row’s version where a line is about a message, else is absent.
• compaction_head stores the winner’s schema_version and the upsert compares (schema_version, compaction_rank, head_id): a newer payload version always wins the key; rank then id decide inside a version. Without it a same-topic bridge write at rank -1 would lose to the v1 head it was made from.
• FamilyHealth becomes TopicHealth: per-version counts inside one topic (rows, compaction heads at the version, each group’s unread and unresolved rows at it); a compacted version’s retire verdict is now answerable, so [0406]‘s permanent refusal ends.

Consequences

• The version is impossible to omit at compile time (missing method), which is [0405]‘s invariant kept and moved earlier. The struct name and its version sit on adjacent lines.
• The bridge [0402] is same-topic: consume v1 rows, re-produce as v2 at CompactionRank -1. [0403], [0404], [0407] unchanged.
• Cost: [0401]‘s per-version isolation (own id space, own retention, own partitions, DROP TABLE retirement) is gone; a v1 row lives until retention drains it, and retiring v1 means “no v1 rows or heads remain”, not dropping a family.
• Rejected: a config-level override of the type’s version (two sources for one fact); an upcaster chain now (ROADMAP Later; the skip predicate is where it would plug in).
• Pre-v1: baseline DDL edited in place; stored rows drop and recreate.