Constant CREATE_LINKS_LOG_INSERT 

pub const CREATE_LINKS_LOG_INSERT: &str = r#"
    CREATE TRIGGER IF NOT EXISTS trg_links_log_insert
    AFTER INSERT ON links
    BEGIN
        INSERT INTO transaction_log (table_name, entity_id, operation, payload, recorded_at, branch_id)
        VALUES ('links',
                NEW.source_id || '|' || NEW.target_id || '|' || NEW.edge_type || '|' || NEW.valid_from,
                'I',
                json_object('v', 1, 'source_id', NEW.source_id, 'target_id', NEW.target_id,
                            'edge_type', NEW.edge_type, 'valid_from', NEW.valid_from,
                            'valid_to', NEW.valid_to, 'weight', NEW.weight,
                            'properties', json(NEW.properties)),
                NEW.recorded_at, NEW.branch_id);
    END;
"#;

The links log, and the entry whose entity_id is composed rather than copied.

source|target|type|valid_from identifies an edge assertion and carries no lineage, which is why branch_id had to become a column of its own rather than a fifth field in that string. Re-keying entity_id was the other option and was rejected: it changes what a log entry identifies, so rows written before the rung would no longer match rows written after it, and the fold would silently split one edge’s history in two.

With the column present, the four folds in temporal::replay — a private module, so the name is plain text rather than a link that would not resolve — partition by (table_name, entity_id, branch_id) and two lineages’ assertions about one edge stay two beliefs. Without it they collapse to whichever has the higher seq_id — no error, no drift report, just one lineage’s belief gone.