scientific: bind Rubin activation to concrete draw-payload digest #515

Scientific finding

Draft #506 now constrains the Rubin activation decision by immutable generator/analysis/evidence authority, source snapshot identity + SHA-256, cutoff, exact validated observation/draw-count points, design envelope, and indicator kind. That closes class-level design generalization, but the positive algorithm still does not bind the activation receipt to the concrete complete-data draw matrix used by the current run.

The private approved-pairing path therefore cannot distinguish two runtime inputs that have the same snapshot, cutoff, observation count, draw count, indicator kind, and approved class-level receipt but different draw values. That violates #505's RED target: an arbitrary caller-constructed matrix must remain distinguishable from the concrete generator output receiving activation authority.

This is not a defect in psychometric_core Rubin arithmetic. Qbar/Ubar/B/T remains reusable numerical truth there. The defect is Validation Evidence / Projection provenance in analysis_engine.

Required repair

Bind RubinProjectionActivationReceiptV1 to the canonical SHA-256 of the exact complete-data draw payload (or an equivalently immutable canonical generated-draw artifact). The activation decision must receive the independently computed runtime draw-payload digest and require exact equality with the receipt before the private approved-pairing path can return Eligible.

Minimum fail-closed behavior:

• malformed or noncanonical draw-payload digest is rejected;
• receipt digest != runtime digest is rejected even when snapshot/cutoff/counts/envelope/indicator all match;
• an exact digest preserves the existing private positive algorithmic branch;
• the digest participates in canonical receipt JSON and receipt SHA-256;
• production PRODUCTION_APPROVED_PAIRINGS remains empty until scientific: establish Rubin loading-uncertainty recovery and coverage evidence #503/test(science): add Rubin loading recovery and coverage acceptance evidence #504 evidence is independently promoted;
• no draw arithmetic is copied into analysis_engine, no LLM judge is introduced, and no mutable locator becomes scientific authority.

This digest prevents payload substitution/replay after receipt creation; by itself it is not proof that an arbitrary caller actually executed the approved generator. If the trusted generator boundary cannot issue the receipt or an immutable generated-draw artifact, generator-execution attestation remains a separate follow-on owner gap rather than something to overclaim here.

RED target

Using the private test-only approved pairing, keep snapshot ID/SHA-256, cutoff, 48 observations, 8 draws, ALR indicator, generator/analysis/evidence identities and design envelope identical. Change only the runtime complete-data draw-payload SHA-256. The exact digest must be Eligible; the substituted digest must be Rejected.

Ownership / fold

Repair in Draft #506, keep ADR 0034 Proposed, and preserve this delta through the eventual ordinary non-force fold into #416. Shared TRACEABILITY/ADR index/product-gap/CHANGELOG current-state authority remains #435's single-writer lane. Do not simple-close this issue until the exact-head repair is verified and inherited by the surviving Analysis Run vehicle.