id: ADR-0042 title: Defer finer memoization until whole-stage measurements justify it status: accepted date: 2026-09-13 deciders: [paul-heyse] level: decision principles: [DM-32, DM-33, DM-56, DM-58] blueprint: [§D14, §14.3, §24.3, §26] review: docs/design_review/reviews/design_review_blueprint-rev5-contracts_2026-09-13.md#6-acceptance-gates evidence: Proposed supersedes: [ADR-0020] superseded-by: null revisit: A measured workload still needs finer dependency tracking after complete per-instance reuse is characterized. verification: just adr-lint; just docs; docs/plans/02-blueprint-revision-5-contracts.md Verification; bash docs/design_review/evidence/blueprint-rev4-2026-09-13/run.sh
ADR-0042: Defer finer memoization until whole-stage measurements justify it
Context
The simpler compiler in review section 8 starts with complete stage keys. ADR-0020 assumed an incomplete per-instance cache and therefore needs a different deferral boundary.
Scope
Amends the granularity and adoption trigger; salsa remains absent from runtime dependencies.
Drivers
Keep one dependency mechanism until a measured workload needs finer tracking. Dependency correctness and cancellation behavior must precede claims of incremental speed.
Options
Adopt salsa or a custom fine-grained dependency engine now: rejected without a measured consumer. Keep the old incomplete tuple: rejected. Refine one existing memo in measured steps: selected.
Outcome
Use the artifact-hash memo at whole-stage granularity first. Evaluate a complete per-instance bundle only after the R-22 measurement trigger. Consider salsa under R-01 only if that measured case requires automatic finer dependency tracking and its cancellation model has a complete boundary contract.
Consequences
Compile latency may initially be higher, but there is one dependency declaration and one cancellation boundary to validate.
Compensating controls
The implementation gates are explicit in plan 02. Document acceptance does not establish runtime behavior. Deferred optimizations remain disabled until their register trigger and conformance evidence are satisfied.
Confirmation
Register R-01 and R-22 track the decisions; plan 02 names the cold/warm and incremental-versus-clean experiments. No speedup is claimed. just adr-lint and just docs check documentation structure only; the plan distinguishes these checks from future behavioral acceptance.
Pros and cons
Deferral avoids a second revision/cancellation model and premature per-instance keys. Whole-stage preparation can cost more; R-22 measures that cost before refinement, and R-01 separately evaluates salsa.
More information
See the governed blueprint sections, the revision-4 review, its reproducible library evidence, the revision-5 contract review, and plan 02.
Status history
- 2026-09-13 — proposed before the revision-5 blueprint amendment; user requested reconciliation of all review findings and library opportunities.
- 2026-09-13 — proposes superseding ADR-0020 for the corrected contract above; accepted arguments remain immutable.
- 2026-09-13 — accepted (revision-5 review, Accept for bounded proposed scope; evidence remains Proposed until the wave-1 implementation review).