id: ADR-0020 title: Defer salsa; the artifact-hash memo is the only memoization mechanism status: superseded date: 2026-09-13 deciders: [paul-heyse] level: decision principles: [DM-58, DM-32, DM-28] blueprint: [§14.3] review: docs/design_review/reviews/design_review_arrow-native-idaes-core-blueprint-rev2_2026-09-13.md#7-principle-findings evidence: Proposed supersedes: [] superseded-by: ADR-0042 revisit: A measurement shows sub-pass granularity finer than per-instance is needed inside P7 verification: The §24.3 compile-time benchmark for P7 on a flowsheet-scale model; register row R-01

ADR-0020: Defer salsa; the artifact-hash memo is the only memoization mechanism

Context

Blueprint §14.3 defers salsa behind the artifact-hash memo with a stated trigger. The first review’s F13 asked whether the two mechanisms duplicate each other; the supporting-library capability map measured that salsa’s backdating and accumulators work as described, which does not change the question.

Scope

Binds the memoization mechanism for phases 0-3. It does not reject salsa permanently; the trigger is a measurement, and the register row carries it.

Drivers

salsa‘s distinctive value is automatic dependency tracking, which duplicates pass_specs’ declared inputs; an undeclared read is a contract violation, not something to track; its unwind-based cancellation cannot cross the Ipopt boundary.

Options

Adopt salsa now — rejected: a second dependency system and a second cancellation model with no measured need (charter DM-58). Adopt it only for P7 — deferred to the same trigger.

Outcome

salsa is not a dependency. Per-instance memoization inside P7 uses the same key shape over (instance_id, template hash, resolved features).

Consequences

If the trigger fires, adoption means reconciling two cancellation models at the Ipopt boundary, which is the part to design before adopting rather than after.

Compensating controls

Register row R-01 carries the trigger, the owner and a phase-gated review date so the deferral does not become an omission.

Confirmation

The §24.3 benchmark measures P7 compile time per instance; the register review reads it monthly.

Pros and cons

Deferring costs a possible re-plumb later; adopting now costs a mechanism nothing has asked for.

More information

Blueprint §14.3 (pass engine); review finding F13; docs/adr/register.md row R-01.

Status history

• 2026-09-13 — accepted with the repository-seeding pull request (backfilled from blueprint revision 3).
• 2026-09-13 — superseded by ADR-0042.