Select the first bounded computational-materials benchmark

[chefmrfrizzle]

What changed

• compares four safe public or synthetic computational-materials benchmark candidates using current primary sources
• adds a transparent weighted scorecard with sourced facts, local measurements, assumptions, estimates, and rights stop-conditions
• recommends exactly one narrow workload: pinned spglib 2.7.0 wurtzite symmetry classification
• defines deterministic logical inputs and discrete outputs, challenge cases, pass/fail rules, a reproducibility budget, failure cases, and required domain expertise
• proposes ADR-0010 without adopting it
• records D-021/D-022 as blocked with accountable owners and required evidence
• records LearningEpisode 0004, including the C/Python lattice-convention counterexample

Why

Prompt 2 requires a legally safe, low-cost workload that can demonstrate deterministic execution, verification, replay, counterexamples, and second-machine feasibility before any implementation begins.

The selected candidate uses a synthetic four-site input and exposes a realistic failure: copying the upstream C lattice directly into the Python row-vector API returns a plausible but incorrect classification.

Scope and governance

This PR contains public research/planning documentation only. It does not implement schemas, adapters, workers, signing, payment, settlement, production infrastructure, or scientific acceptance policy.

Disclosure: @chefmrfrizzle and requested reviewer @11BUSD share one human operator. Any @11BUSD approval is separate-account workflow evidence under ADR-0009, not independent-person or domain-expert assurance. ADR-0010 remains proposed; D-018, D-019, D-021, and D-022 remain blocked.

Validation

• ./tests/test_repository_baseline.sh
• BASE_SHA=origin/main HEAD_SHA=HEAD ./scripts/check_repository_baseline.sh
• git diff --check
• all 26 primary-source links checked; 25 returned HTTP 200 and one publisher DOI endpoint returned HTTP 403 to the automated client while remaining a valid DOI
• local discovery run: macOS arm64, CPython 3.9.6, spglib 2.7.0, NumPy 2.0.2; 1,000 repeated classifications produced one selected result

Review focus

• scientific defensibility of the claim boundary and perturbation cases
• whether sourced facts, measurements, assumptions, and estimates remain clearly separated
• whether dataset/code licensing conclusions are appropriately conservative
• whether D-021/D-022 block implementation until domain and second-machine evidence exist