Improve community launch readiness

[FU-max-boop]

What changed

• Rework the README around a 30-second provider-free check, honest upstream impact, and a clear comparison with the SDK's public ScriptedModel.
• Add a real Runner function-tool demo for both run and run_streamed, with exactly-once and script-consumption tests.
• Add a five-minute RunState case study that traces the released failure through merged upstream fixes #4413 and #4414.
• Add contribution, security, issue, and pull-request entry points plus Discussions links.
• Improve package discovery metadata and add reusable social-preview assets.

Why

AgentRunProof already has verifiable upstream diagnostic impact, but the repository's first screen and contribution path were too research-oriented for a new user. This change makes the value reproducible in 30 seconds while keeping the recognition boundary explicit: the upstream fixes cited the findings, but OpenAI has not adopted or endorsed the package.

Verification

• openai-agents==0.20.0: 168 passed, 1 expected skip
• installed wheel with openai-agents==0.21.0: 169 passed
• ruff check .
• ruff format --check .
• mypy src
• fresh wheel and sdist build; twine check --strict passed
• clean wheel and sdist installs passed pip check, CLI probe, and certificate verification
• README Markdown, issue-template YAML, SVG, and 1280x640 PNG validated

Compatibility and evidence

No runtime behavior, certificate schema, canonical evidence, supported SDK range, or immutable release artifact changes. The package version remains 0.2.0; these changes are recorded under Unreleased.