Retracted means retracted

[WaylandYang]

Fixes #202.

"Data is wrong" only marked the violation resolved; the fact stayed in the graph, and the next check hit the resolved row and stayed silent — the ledger said retracted, the graph disagreed, and the queue never mentioned it again.

The decision names a fact and retracts it. POST …/violations/{id} takes fact_id for fact_retracted; single-fact kinds (self loop, signature, derived contradiction) pick it themselves, and pick_retraction accepts only a fact the violation lists (left, right or the cycle path). reasoning::retract_from_violation goes through reject_fact — invalidated_at, evidence untouched, ledger snapshot in the audit log — then marks the row resolved, and the handler reruns the check so any other violation that fact was part of clears too. The cards for asymmetry, functional and cycle show a Retract button per fact; cycle cards now list the path (open_violations returns path as id + text).

The queue stays honest. When run recomputes a violation whose row is resolved with a resolution that promised a change (fact_retracted, fact_closed, axiom_relaxed), the row is reopened — the promise did not hold. accepted keeps its silence. The report counts reopened.

Test: a_retraction_leaves_the_graph — retract one side of an asymmetry, the other stays and the rerun is quiet; a fact outside the violation is refused; a self loop needs no choice; axiom_relaxed with the axiom still declared reopens, accepted does not.

Verified in the isolated environment: an asymmetric pair shows two Retract buttons, one click retracts that fact, the card leaves the queue, the ledger and audit log agree.

0002 carries a revision note.

🤖 Generated with Claude Code