Prepare the v0.3 release evidence profile

[FU-max-boop]

Summary

• add the v0.3.0 release profile and move canonical comparison generation to a fresh v3 directory
• continue validating immutable v1/v2 evidence while requiring the v3 evidence commit to be the direct child of its clean source commit
• finalize truthful release-prep metadata without claiming that v0.3.0 is already published

Release impact

This PR does not publish, tag, or add canonical v3 evidence. After merge, the clean source commit must pass the full hosted matrix. A separate manual Linux run will then generate the v3 bundle from the audited OpenAI Agents fix commit. The follow-up evidence PR must add only five certificates and bundle.json.

The v3 directory remains comparison schema v1 and covers the existing 0.20 causal ladder. SDK 0.22 support is established by the exact 15-cell packaged-wheel matrix, not by that comparison bundle.

Verification

• ruff check .
• ruff format --check .
• mypy src
• pytest -q: 184 passed, 1 expected skip on local SDK 0.20
• focused workflow tests: 30 passed
• python -m build
• twine check --strict: wheel and sdist passed
• fresh CPython 3.12 wheel and sdist installs against openai-agents==0.22.0: pip check, CLI probe, and certificate validation passed

No runtime source, schema, requirements lock, generator script, or existing evidence file changed.