one_command_setup assessor should check CLAUDE.md/AGENTS.md, not just README

[jwm4]

Summary

The one_command_setup assessor (assessors/structure.py) only searches README.md for setup commands via _find_setup_command() and _is_setup_prominent(). It does not check CLAUDE.md or AGENTS.md.

For agent-readiness, CLAUDE.md is arguably the more important location. An agent reads CLAUDE.md first, so a setup command documented there is immediately actionable. A repo like mlflow/mlflow has its dev setup command (uv run bash dev/run-dev-server.sh) documented in CLAUDE.md but not prominently in their README, and scores 60/100 as a result.

Proposal

Have _find_setup_command() and _is_setup_prominent() also search CLAUDE.md and AGENTS.md. A setup command found in an agent context file should earn at least the same credit as one found in README.

Context

Discovered while assessing mlflow/mlflow (score: 60.4, Silver). This was one of the attributes where the score didn't reflect the repo's actual agent-readiness.

Posted by Bill Murdock with assistance from Claude Code.

[unitedideas]

The README-only scoring gap looks right. I would split it into two checks:

1. Discovery: does the repo expose setup commands in README.md, AGENTS.md, or CLAUDE.md?
2. Validation: does the command actually run from a clean checkout or CI job?

The second check matters because instruction files often contain commands that are right for an agent but wrong for a fresh machine: local path assumptions, missing env refs, or prior bootstrap state.

This AGENTS.md / CLAUDE.md migration checklist covers some of the edge cases:
https://bringyour.ai/agents-md-claude-md

The import-audit framing is similar: treat agent context files as executable setup hints, then prove the setup path before assigning readiness credit:
https://bringyour.ai/codex-import-audit