Internal working engineering record - kept for reference, not a user-facing guide.

Typed stats are the Ingest Stage contract; the DB is the inter-stage data contract

axctl will model each Ingest Stage as a co-located StageDef whose run returns a typed BaseStageStats-extending value carrying durationMs, summary, and stage-specific counts. Stages do not pass in-memory row collections to downstream stages; downstream stages read evidence back from SurrealDB. Deps express ordering, not data flow.

The deletion test confirms this: most stages share ordering, not row payloads. Stages like derive-opportunities and derive-retro-proposals already read cross-table evidence (proposals, signals, friction events, edges) that Surreal joins better than in-memory plumbing. Passing structured rows through the Ingest Pipeline would create one-adapter pretend-seams without a real second consumer.

Stage keys and tags are themselves co-located Schema.Literals - each stage file exports SignalsKey = Schema.Literal("signals"); a central registry composes the union via Schema.Union. Tags follow the same per-tag-literal pattern with a central union. JSDoc {@link} references between keys give IDE-navigable dependency graphs and a place to scribble TODOs about not-yet-wired stages.

Consequence: adding a stage edits one file plus a tiny registry import; typed stats survive end-to-end through Effect; pipeline reasoning stays focused on ordering and concurrency rather than buffer management.