Add AgentRunProof to eval frameworks and harnesses

[FU-max-boop]

What changed

Adds one entry to §5a · Eval frameworks & harnesses.

Why it belongs

AgentRunProof is a provider-free runtime-regression harness for the OpenAI Agents SDK. It drives the real Runner with deterministic scripts, compares run and run_streamed behavior, checks RunState resume invariants, and emits content-addressed records.

This is deliberately not presented as a model-quality evaluator. Its scope is SDK runtime correctness.

The public case study documents the evidence chain around two merged, member-authored upstream fixes:

• fix: isolate RunState checkpoint tool decisions openai/openai-agents-python#4413
• fix: resume recursive agent tool approvals openai/openai-agents-python#4414

The entry explicitly states that this is not OpenAI adoption or endorsement.

Evidence of outside use and review

• A LoopGain maintainer independently ran the proposed integration branch while it still installed AgentRunProof and reported 208 passed, 5 skipped: test(openai-agents): exercise the real Runner offline loopgain-ai/loopgain#14 (comment). He then requested the narrower SDK-native fixture, so the merged default branch does not retain AgentRunProof. This is an external trial and technical review, not adoption.
• An unaffiliated author included AgentRunProof in a third-party survey of auditable-agent infrastructure: https://github.com/UniqueClouds/marginalia/blob/main/docs/entries/011-auditable-agents-atlas.en.md#L248. This is a write-up/recognition signal, not execution evidence.

Disclosure and maturity

I maintain AgentRunProof. It is still new, has no claimed default-branch package adoption, and remains unproven beyond the public evidence and external trial above. The inclusion case is reproducible runtime evidence, upstream diagnostic impact, and the disclosed outside review—not stars or self-reported benchmark results.

Verification

• Checked README, MENTIONS, notes, and all-state issues/PRs for the project name and canonical URL; no duplicate found.
• Verified the repository, case study, upstream PRs, LoopGain report, and third-party write-up links.
• Rebased onto current main; git diff --check passes.
• Changed only README.md with one annotated entry.