M5: v0.5 evals

[hazeliscoding]

What an asset does once it's active, measured on the real Claude Code and Codex in a sealed home, and whether a change made it better. The plan and its decisions are in ROADMAP.md under M5 and "Evals (v0.5)".

• axm eval run: each asset's eval cases, each session in a sealed home with only the user's logins and chosen model, in its own git copy of the case's repo with the asset installed. Checks decide each run; an optional rubric is graded by a model as model judgment, apart from the checks.
• axm eval compare: a baseline at a git ref, or no asset, against the working tree, with sessions alternating, a side-by-side report, and a saved baseline reused while it still matches.
• axm conflicts <path> --judge: contradictions between the instructions a file gets, kept only when the model quotes both passages from the files.
• Behavioral evals for the four starter assets, and agent-asset-authoring 0.2.0, whose change the first compare measured.
• Codex on Windows reads the whole disk and starts its sandbox slowly; the report names paths a run touched outside its copy and each case's slowest call, and docs/cli.md states the limits.
• Write-up: docs/did-the-change-make-the-skill-better.md.