Retrospective packet

Use DOC, data.kind=retro. Link task, session/run, review, checks and FDK.

Observation	Evidence	Cause/confidence	Proposed change	Owner	Next-run measure
Repeated work after reconnect	Duplicate operation trace	Missing identity, high	Retain/replay command key	Task implementer	One local revision
Review lacked native proof	Review report gap	Acceptance too vague, medium	Add real tool-use observation	Spec owner	Evidence linked in report

Keep: practices supported by successful evidence. Change: small owned actions, not broad slogans. Unknown: missing events or untested causes. Follow-up: TSK/RDT/DOC/FDK/MEM as appropriate, with proposal/application status explicit.

Do not infer a personal failing or universal rule from one bad run. On the next relevant task check whether the accepted change actually helped. Retain the superseded lesson if it did not.