[kvlonge]

Why

Setting up a shared ClickHouse Recall store still needed the database created by hand, and the docs and agent skill presented running sqb recall schema --ddl yourself as the setup route. After the version row expired, a store with previous-version views (same columns) could be recorded as current without its views being replaced. Orchestrator context was stored as a Python repr, so it could not be queried.

Changes

• A writer allowed to run DDL creates a missing Recall database (CREATE DATABASE IF NOT EXISTS) and says so once, even if later setup fails or is retried, followed by a plain success line. Existing databases are untouched. A writer that may not gets the exact statements; sqb recall schema --ddl prints them too.
• Views whose definitions set join_use_nulls count as the previous version whatever the version row says: writers with DDL replace them, restricted writers warn. Newer stores are never downgraded.
• invocations.external_context is stored as JSON.
• Docs and skill: Recall creates and upgrades its own store, with tested writer grants and a note for self-managed clusters; --ddl only for restricted writers. Also store selection, recording, the default format, checking values before filtering, common queries and statements.query_id.

Verification

• ClickHouse E2E: database created with notice (all rights, documented grants); existing database not re-created; refused writer gets statements and writes after they run; database created but tables refused reports both; documented grants replace older views and add a missing column; older views replaced or warned; JSON job query.
• Writer test: the creation notice survives refused and retried setup, exactly once.
• Integration test runs the documented common queries from the Recall page.
• Existing ClickHouse Recall E2E and Recall integration tests, sqlbuild-recall tests, make check-ci and the hygiene scanner pass.