Four serialized Swift Testing suites pit the shipped implementations from the #644-#665 wave against verbatim copies of the code they replaced: the untracked line-count scanner vs the Data.reduce reader, normalize vs the regex-always formulation, the worktree directory index vs the per-row PathPolicy scan, and warm TranscriptFragmentCache scoring vs cold re-parsing. Each suite asserts output equivalence before timing, and asserts a floor ratio measured interleaved so machine speed and load cancel out. Ratios whose slow side is a C call or filesystem I/O assert in every build mode; the two Swift-vs-Swift ratios (escape guard, dense-input hybrid fallback) collapse under -Onone and assert only in optimized builds, detected at runtime through an assert side effect. Release test builds surfaced two pre-existing snags: tests exercising Debug-only members (CommandIconMap.debugAllEntries, CLISocketServer.debugFileDescriptors) now compile away outside Debug, and GhosttyRuntime declares @MainActor explicitly because whole-module swiftmodule deserialization loses default-isolation inference for its isolated deinit. Claude-Session: https://claude.ai/code/session_01TWSNesKV4dk8HLKmcNEJjS