Add v0.3 upstream comparison evidence

[FU-max-boop]

Summary

• commit the fresh v3 upstream comparison bundle generated on canonical Linux / CPython 3.12.13
• bind the bundle to source commit 22c264a and the AgentRunProof 0.3.0 candidate wheel
• preserve the five-run OpenAI Agents 0.20 causal ladder while keeping v1 and v2 immutable

Verification

• canonical workflow run 32455709100: success, including all 15 packaged-wheel Python/SDK cells
• bundle validates as sha256:f2025a7e71937e4fb413287cf166d23ac74bb6ae7f02a7bff8b1f93f61b044bc
• harness wheel matches the distributions artifact byte-for-byte
• local suite: 184 passed, 1 expected skip
• this PR is the direct child of the recorded source commit and adds only six JSON evidence files

Scope boundary

This v3 comparison refreshes the OpenAI Agents 0.20 RunState causal ladder. OpenAI Agents 0.22 compatibility is supported separately by the hosted 15-cell packaged-wheel matrix; this bundle does not claim to evaluate 0.22.