retrospectlisted

# Retrospect Turn completed-cycle evidence into a better next cycle: benchmark, diagnose, propose, verify, monitor. No retrospective theatre or one dated log per run. Scale depth to risk/friction: clean routine cycles may return `no change`; substantial runs, escaped defects and repeated human corrections need the full pass. ## Evidence boundary Start from approved spec, run receipts, checks/evals, review/repair history, human corrections, escaped defects, production observations, routing failures, resource use and retained artifacts. Use only authorised sources. Mark absent dimensions `unknown`; never infer success from a clean final answer. For substantial+ or repeated cycles, create `RETROSPECT.json` from the [template](templates/RETROSPECT.template.json) and validate with `scripts/validate_retrospect.py`. It is evidence, not diary/project truth. ## Review dimensions - outcome and acceptance-criteria success; - trajectory compliance, safety and escaped defects; - human attention/corrections/rework, gate latency, unnecessary interruption and blocked time; - test/eval and reviewer effectiveness; - orchestration, model routing, delegation, tool reliability; - context size, compaction, handoff and artifact hygiene; - skill triggering, usefulness, overlap and missing capability; - documentation, project memory, canonical-state freshness; - cost and latency when receipts contain trustworthy measures. ## Flywheel 1. **Benchmark** against declared acceptance criteria/eva