Sync both -semantic rubric agents to v2.1 source rubric fields (RAI, Data Protection & Compliance, composition)

[realmarcin]

Context

Follow-up from review of #154 (rubric-review branch).

Problem

The embedded rubric spec inside both -semantic agent files predates the v2.1 source rubric (data/rubric/rubric10.txt, "Last Updated: 2026-01-12"). The agents are missing the schema fields that the v2.1 rubric now references — and that D4D-core surfaces with explicit ontology grounding (schema.org / RAI / EVI / DCAT).

Confirmed absent (grep count = 0) from both d4d-rubric10-semantic.md and d4d-rubric20-semantic.md:

v2.1 / core field	Where it belongs
data_protection_impacts, reidentification_risk, hipaa_compliant, other_compliance, governance_committee_contact	Data Protection & Compliance (rubric10 Elem 2/4)
known_biases via BiasTypeEnum ("RAI-aligned"), future_use_impacts (rai:dataSocialImpact)	RAI / Croissant alignment (Elem 9)
missing_data_documentation, annotation_analyses, machine_annotation_tools, imputation_protocols, raw_data_sources	Composition / Technical (Elem 5/6/8)
is_data_split, is_subpopulation	Composition (Elem 5)

Why it matters

D4D-core's entire purpose is RAI/EVI/DCAT/schema.org grounding, yet the rubric gives no credit for those mapped fields. #154 touched these files and was the natural place to sync them, but didn't.

Proposed fix

Bring the "Rubric Specification" sections of both -semantic agents into sync with data/rubric/rubric10.txt / rubric20.txt v2.1 sub-element names and field lists (e.g. "Biases Categorized Using Standard Taxonomy (RAI-aligned)", "IRB or Ethics Review and Data Protection Impact", "Privacy Protections and Re-identification Risk Assessment", "Provenance, Source Derivation, and Raw Data Sources", "Preprocessing, Cleaning, Labeling, and Annotation Quality", etc.). Consider whether the non-semantic d4d-rubric10.md / d4d-rubric20.md should be synced in the same pass (they lag too).

Scope

• .claude/agents/d4d-rubric10-semantic.md
• .claude/agents/d4d-rubric20-semantic.md
• (consider) .claude/agents/d4d-rubric10.md, .claude/agents/d4d-rubric20.md

[realmarcin]

Adversarial review (Codex) verdict: KEEP — blocking. Scope expanded (absorbs Codex "Finding 4").

Confirmed v2.1 drift: data/rubric/rubric20.txt:273-280 and :493-500 (incl. Q20 RAI) carry fields that d4d-rubric20-semantic.md:197-200 and :386-397 omit.

Add to this issue's scope a regression introduced by PR #154 itself: the new ethics-plausibility prose re-anchors checks to section-name headings ("Licensing & Governance", "Data Protection & Compliance", "Human Subjects") that are not field names a model sees in a datasheet — D4D-core exposes license_and_use_terms, data_protection_impacts, human_subject_research, etc. A YAML with correct field-level data but no matching prose headings can be misread. Relatedly, Q8's field list (.claude/agents/d4d-rubric20-semantic.md:199) omits data_protection_impacts and reidentification_risk even though the v2.1 source includes them.

Fix: anchor the plausibility checks to exact field names — license_and_use_terms, ip_restrictions, regulatory_restrictions, data_protection_impacts, reidentification_risk, human_subject_research, informed_consent, participant_privacy (+ Core equivalents) — and add the missing v2.1 privacy fields to Q8/Q9.