he-learn

name	he-learn
description	Turn a verified process failure into narrow durable prevention without blocking normal product delivery.

Hard Eng Learn

Contract

• Input = explicit $he learn evidence OR verified stage trigger.
• Output = accepted learning record + asynchronously assigned prevention, or exact non-candidate.
• Lifecycle = unchanged; overlay only.
• Load workflow.md before classification or repair.

Ownership

• Failure classification + prevention owner + follow-through = this skill.
• Recurrence/root-cause proof = $repeated-failure-learning.
• Prevention implementation = destination owner under its normal direct/$he route.
• Product implementation = current stage owner; learning does not seize it.

Invariants

• Trigger = verified recurrence OR user correction OR false gate OR systemic protected-boundary gap OR repeated manual waste.
• One-off implementation finding = current build loop; learning candidate = forbidden.
• Record fact once = failure + evidence + root class + prevention owner + required proof.
• Narrowest durable owner = invariant/schema/code/test → deterministic rule/tool/CI → skill/route → docs/runbook.
• Repair = asynchronous + non-blocking for normal product delivery.
• Block/pause only when continued product work risks security/privacy/data loss/irreversible action or another protected boundary.
• Cross-repository repair = destination-owned normal flow; routine source PLAN pause/writer lease/nested lifecycle = forbidden.
• Open learning work does not block green|shipped unless protected-boundary risk remains.

Complete

• Candidate validity = evidence-backed.
• Prevention owner + proof = explicit.
• Applied prevention = regression-proven; deferred prevention = tracked without blocking unrelated delivery.