rubric10-semantic: add conditional applicability to match rubric20 (core datasheets penalized for by-design omissions)

[realmarcin]

Context

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

Problem

#154 added conditional applicability ("Always report results… but only score if identified elsewhere") only to d4d-rubric20-semantic.md. d4d-rubric10-semantic.md was left with strict binary present/absent scoring.

This asymmetry directly hurts D4D-core evaluation. D4D_Core.yaml intentionally excludes fields:

"Fields not in the exchange layer mapping (e.g., file_collections, relationships, splits, consent workflow) are intentionally excluded; use the full data_sheets_schema.yaml for those."

So rubric10-semantic will penalize claudecode_agent_core datasheets (the new *_core_evaluation.json runs) for absences that are by design — while rubric20-semantic now degrades gracefully for the same datasheets.

Proposed fix

Either:

• (a) Add the same conditional-applicability treatment to rubric10-semantic sub-elements (report but only score when the relevant context is present elsewhere in the datasheet), or
• (b) Add an explicit "absent-by-design ≠ fail when evaluating a D4D-core datasheet" stance, with a documented list of core-excluded fields that should not be penalized.

Depends on / coordinates with the N/A-encoding convention (companion issue) so rubric10 scores stay comparable.

Scope

• .claude/agents/d4d-rubric10-semantic.md

[realmarcin]

Adversarial review (Codex) verdict: KEEP — blocking.

Confirmed: src/data_sheets_schema/schema/D4D_Core.yaml:124-126 documents that core intentionally excludes fields (file_collections, relationships, splits, consent workflow), yet d4d-rubric10-semantic.md retains strict binary present/absent with no applicability handling — so core datasheets are penalized for by-design omissions. Should be designed together with the #155/#161 applicability+N/A mechanism so rubric10 and rubric20 behave consistently.