Failures, document_gap, and quality checks

Failures are auditable terminal or blocked states. Preserve phase, time, command/response, error text, and the registry writes for every failure; do not hide the cause with retries.

Phase mapping

Execution failures are plan_failed, submit_failed, upstream_failed, finalize_failed, and downstream_failed. Authoring failures are draft_failed, build_failed, test_failed, publish_failed, register_failed, and verification_failed. A failure phase cannot jump back to success; retrying means opening a new, independent workspace.

document_gap

When the AI docs are unavailable (including the current 404 endpoint) or lack information needed for a decision, record:

• topic: the decision under review;
• attempted_document_ids: document IDs already tried;
• missing_information: missing fields, rules, or examples;
• blocking_reason: why it is unsafe to continue.

Also set detected: true, workflow.phase: document_gap, and status: blocked, then stop. Do not fill the gap with an old page, memory, or implicit default. Once an authoritative document is available, the user decides whether to open a new workspace.

Release checks

1. Both registries have the correct lane, schema version, path root, and document source with no cross-lane fields.
2. The execution plan is confirmed before Job creation; authoring build, publish, and register each have their own confirmation.
3. tested means static validation passed: Tools and Tool Versions check config, bilingual summaries, command fields, and resources; Flows check exact node Tool definitions, ports, required inputs, parameters, DAG, method evidence, and sourceGraph read-back.
4. Every ID came from fkit and passed list/config/summary/inspect read-back.
5. Downstream prepares an exact NAME:VERSION, verifies its bilingual summary and empty config command, and then pulls it. Network is none; stage and task reproducibility are read-only; scripts, image identity, logs, and results are traceable.