Harden agent compatibility scoring

[garethpaul]

Summary

• pin and wrap the deterministic scanner with executable version, path, output-shape, and repository-classification checks
• make a fail-closed synthesizer the sole authority for score arithmetic, degraded states, target identity, evidence, and isolated execution provenance
• tighten specialist safety boundaries and structured handoffs, including separate temporary copies for stateful reviews
• add contract tests, Cloudflare Worker and genuine CLI fixtures, symlink-install regressions, release metadata, and CI coverage

Validation

• npx -y node@20 --test agent-compatibility/test/*.test.mjs (36 passed; one Windows-only hosted-runner assertion skipped on Linux)
• node scripts/validate-plugins.mjs
• Prettier 3.6.2 check across every changed file
• git diff --check
• reproduced scanner 0.1.7 misclassifying a real Cloudflare Worker as a CLI; the new guard correctly returned unreliable and excluded that deterministic score
• added shell-free Windows scanner invocation plus Ubuntu and Windows CI matrix coverage after review feedback
• three independent critic rounds; final verdict 9.4/10 with no blocking findings

Limitation

Full end-to-end orchestration inside an installed Cursor runtime was not available in this environment. The documented symlink-install boundary is covered by executable regression tests.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before autofix could start.}

^{Reviewed by Cursor Bugbot for commit 0a36372. Configure here.}