Paired fixtures. Each sample is a folder with legacy/ (a .NET Framework 4.8 solution,
old-style csproj) and modern/ (a .NET 10 solution), plus, from M3-003, expected.sarif.json (the
snapshot the integration tests assert against) and a README.md stating which verdicts
the sample is designed to produce. Fifteen samples pair other runtimes (ADR 0040, P2-055, P2-113, P1-029, P2-119, P2-143, P2-145, P2-117, P2-114, P2-118, P2-165, P2-164), with an
SDK-style project on both sides: same-runtime-cleanup, cleanup-modern-syntax,
cleanup-extract-method, async-disposal, webapi-inherited-route and extern-import (.NET 10 on both), version-bump, params-span-overloads and generic-enum-members (.NET 8 against .NET 10) and runtime-row-framework-only-change, generated-regex-unedited and generated-regex-edited
(.NET 8 against .NET 9) and runtime-row-sync-position, extern-import-runtime and append-interpolated-string (.NET 5 against .NET 6). The two
cleanup-* samples (P2-049) record what equiv says today about behaviour-preserving cleanups, so
some of their expected verdicts are known gaps, each with an owning ticket in the sample's README.
Samples are deliberately tiny: a handful of methods each, one concept per sample. They are the executable specification of the tool. The initial set is defined in ticket M1-001 and grows with every frontend feature.