id: ADR-0025 title: Defer LogicalPlan::Extension rule nodes; derivations already answer the attribution question status: accepted date: 2026-09-13 deciders: [paul-heyse] level: decision principles: [DM-58, DM-46, DM-49] blueprint: [§14.2] review: not-required: recorded in the second review’s observations as an adopt-candidate, not as a finding evidence: Proposed supersedes: [] superseded-by: null revisit: The §22.4 diff report or agent tooling needs per-rule attribution inside a plan rendering verification: Register row R-02; the §22.4 diff-report acceptance test when it is written

ADR-0025: Defer LogicalPlan::Extension rule nodes; derivations already answer the attribution question

Context

Blueprint §14.2 rule 8 defers wrapping each rule body in a LogicalPlan::Extension node carrying rule_id. The DataFusion capability map measured that such a node survives the optimizer intact (PROBE B), so the deferral is a cost decision, not a feasibility one.

Scope

Binds whether rule attribution lives in the plan. It does not affect provenance.derivations or rules_fired, which answer “which rule produced this row” today.

Drivers

The mechanism costs the 14-method UserDefinedLogicalNode contract and an ExtensionPlanner; the question it answers is already answered at the data level (charter DM-58).

Options

Adopt now — rejected: machinery ahead of a demonstrated consumer. Never adopt — rejected: EXPLAIN-level attribution is genuinely the better answer once a diff report renders plans.

Outcome

Not adopted. Derivations and the optimizer observer’s rules_fired list are the attribution mechanism for phases 0-3.

Consequences

If the trigger fires, the node must be added for every rule body at once, because partial attribution in an EXPLAIN is worse than none.

Compensating controls

Register row R-02 carries the trigger and a phase-gated review date.

Confirmation

The observer closures of Analyzer::execute_and_check and Optimizer::optimize already record which rules fired on each plan, which is recorded in the pass record (ADR-0019).

Pros and cons

Deferring costs a later retrofit across all rule bodies; adopting now costs a 14-method contract with no consumer.

More information

Blueprint §14.2 rule 8; the second review’s observations; docs/adr/register.md row R-02.

Status history

• 2026-09-13 — accepted with the repository-seeding pull request (backfilled from blueprint revision 3).