[kvlonge]

Why

A SQLBuild command failing under Dagster did not name the failed audits, the selection flooded the run log and metadata, and failed checks were debug-level events. Recall could not tell a local sqb debug from a failed scheduled run, finding a job's runs meant parsing external_context, and an ambiguous ID prefix gave no help.

Changes

• Dagster: the failure description starts with up to 10 failed checks and assets (name, model, failing rows), then "and N more"; metadata gains failed_checks and failed_check_count.
• Dagster, breaking for metadata readers: the failure metadata key selection is replaced by selection_count plus selector_file; sensors that read selection must switch.
• Dagster: the selection logs as a count at INFO, in full at DEBUG; failed checks log at ERROR (WARNING if warn).
• Recall: invocations records origin (integration, ci from CI variables, or local) and command_mode for sqb debug. Additive columns, schema version 4, NULL on older rows.
• Recall: the invocations view exposes integration, integration_job, integration_run_id and integration_retry on SQLite and ClickHouse; older non-JSON context reads as NULL.
• Recall: output shows 8-character invocation IDs; an ambiguous prefix lists up to ten matches.
• Rules: a plain @rel.keys group read in a SQL macro body no longer trips SQBRSQL028/029; real findings beside it still report.
• Docs, skill and sqb recall schema describe all of the above.

Verification

• SQLite CLI tests for origin, mode, integration columns (with a legacy repr row) and prefixes; a Dagster E2E with a failing not_null audit checks the description, metadata, ERROR log and origin; a DuckDB E2E for macro body rules.
• ClickHouse E2E cases for schema version 4 and the integration columns (CI).
• Rust unit tests for the summary, origin, mode, candidates and slot reads.
• make test (before the rules fix), make check-ci, make check-e2e-shards and the hygiene scan pass.