Software Factory Flow Metrics

Deterministic quality, reliability, and flow metrics for a @swamp/software-factory work item — time-to-terminal, per-stage durations, entry counts and time-to-gate, dispatch attempts, failed/parked stage, human touches, per-gate rejection counts, human cycle-limit overrides and the stages they unblocked, patch cycles, delivery mode, and terminal outcome — with a cross-run aggregate, every number traceable to the journal or artifact record it came from, rendered statically from recorded run data with no LLM involved.

Gate-differentiated rejection counts and cycle-override/exhaustion visibility.

Run-level:

• rejectionsByGate: rejections keyed by the gateId each 'rejected' journal event carried, so you can see which gate is actually costing a run. Read from the journal, not the approval records, which are last-write-wins per gate id and lose a rejection that was later approved. Null-prototyped; a gateId-less event buckets to '(unknown)' so counts always sum to the flat rejections total.
• cycleOverrides: human cycle-limit grants ('cycle-override:' approvals), each with the granting stage and the gate id. These differ in practice: a grant for cycle-override:review is commonly recorded while the run sits in a preparation stage.
• cycleExhaustions: stages driven past their limit, keyed on the overridden stage. Documented honestly as an inference: the engine throws on a cycle-limit breach without journaling it, so a run sitting at an un-overridden limit is not recoverable from recorded data.

Aggregate: totalRejectionsByGate (merged across runs) and totalCycleOverrides.

Renderer rows for all of it. README documents each field's JSON shape plus verified consumer queries against the materialized report data, and the latest-version-only caveat on swamp data query.