Add agents.txt v1.0 at repo root

[barneywohl]

Most of the agents I run (and the one filing this PR) reach for anthropic on day one, so adding an agents.txt here felt worth a try. Happy to sign the Anthropic CLA if it isn't auto-detected on this PR — I noticed prior external contributors call it out in their PR bodies, so flagging up front.

This is complementary to CONTRIBUTING.md, not a replacement. agents.txt is a small, machine-readable v1.0 standard (lineage: robots.txt → sitemap.xml → llms.txt → agents.txt) that tells autonomous agents what they can and can't do in a repo before they open a PR. Spec: https://github.com/barneywohl/agentpress/blob/main/docs/AGENTSTXT_SPEC.md

What this adds

• A single agents.txt at the repo root
• Steers agents toward examples/, tests/, and docs — the surfaces your CONTRIBUTING.md already calls out as generator-safe
• Marks generated paths (src/anthropic/_client.py, _models.py, types/**, resources/**), .stats.yml, and release-please config as requires_human_approval
• Prohibits publishing, editing the OpenAPI surface directly, racing the Stainless release PR pipeline, and bypassing CODEOWNERS
• Points test_command and lint_command at the existing ./scripts/test and ./scripts/lint

What this doesn't do

• No CI changes, no workflow edits, no new dependencies
• No code changes anywhere else
• Not a CLA-style or licensing instrument — purely a contribution-policy hint file for agents
• Doesn't try to encode anything an agent couldn't already infer from CONTRIBUTING.md; it just makes it parseable

Caveat I want to be upfront about

This SDK is Stainless-auto-generated, so the human-touchable surface is genuinely small — mostly examples/, tests/, and src/anthropic/lib/. The file reflects that honestly rather than pretending the whole repo is open for agent edits. If the maintainers would rather agents stay out entirely, that's a one-line change to [allowed_actions] and I'm glad to make it.

Standard is brand new (shipped this week), so there's no expectation of adoption — happy to close this if it isn't a fit, or to revise based on whatever the SDK team would prefer agents do or not do here.