eval: #1258 hard negatives for adjudication and a direction metric for edge typing

[jasonssdev]

Summary

The two prerequisites #1269 lists before any bake-off run. Only evals/ changes; no engine code, and no model was run, so no recall or accuracy numbers are reported.

A. #1258-shaped hard negatives in the identity adjudication fixture (evals/adjudication/)

• New probe class procedure-about (3 pairs): a Concept against a Procedure about it (e.g. "Lumen Toolkit" vs "Installing Lumen Toolkit"). Each procedure carries type_alternative="Concept" so the production cross-type bridge nominates it.
• New probe class ui-component (2 pairs): a framework against its own UI, both Concepts (e.g. "Tessera" vs "Tessera Web UI") — the part-whole exclusion the rubric already states.
• Every new pair is expected different; all content is synthetic. Stored arms predate these classes and are not comparable with runs that include them; the README says so.

B. A direction metric for edge_typing (evals/edge_typing/)

• direction_pairs() pairs each forward edge with an asymmetric label to its reversed twin whose trap_type is that label: 9 pairs covering 18 of 29 labelled edges. Symmetric labels, abstentions and forward edges with no reversed probe are excluded, not counted as passes.
• score_direction() classifies each pair-run as discriminated, blind (same asymmetric type both ways) or abstained (no correct forward claim). The existing trap-hit count only looks at reversed probes, so it cannot tell a direction-blind model from one that is right forwards.
• The report prints "Defined on 18 of 29 labelled edges" and each outcome as n of total pair-runs; counts go into runs-*.json.

Judgement calls worth a look: the new classes are small (rates move in steps of 1/3 and 1/2); they are negatives only, relying on the existing *-same controls; and a wrong asymmetric forward type lands in abstained rather than being charged to direction.

Related issue

Refs #1269

Type of change

• feat — new feature
• fix — bug fix
• docs — documentation only
• refactor — no behavior change
• test — tests only
• chore / ci — tooling, build, or CI
• Breaking change

How was this tested?

• adjudication --self-test: 27 labelled pairs across 11 probe classes materialize into exactly that many candidate groups.
• edge_typing --self-test: 11/11, including oracle, direction-blind and all-related_to synthetic models with known expected counts. Mutation: flipping the reversed-answer comparison in score_direction turns it red (9/11).
• ruff check, ruff format --check, mypy ., pytest --cov (96.46%), evals/run_self_tests.py (46/46) pass locally.

Checklist

• My commits follow Conventional Commits.
• I added or updated tests for the change.
• I updated docs where behavior, interfaces, or the knowledge model changed.
• Lint, format, type check, and tests pass locally (ruff, mypy, pytest).
• Output remains OKF-conformant and derived stores stay reconstructible from the bundle + sources.
• The change is consistent with the project's guiding principles (local-first, provenance, freshness, human-in-the-loop).