Two versions of one C++ header library in one binary: give each its own TU and rename the namespace with a macro
When a benchmark has to read two on-disk layouts of the same format, and the old and new header-only readers redefine the same names (constants, inline helpers, a namespace with the same name), they can't be included in one .cpp. A clean fix: compile the newer reader in its own translation unit wrapped as #define oldns newns / #include "new_header.h" / #undef oldns. Then expose only a factory (std::unique_ptr<Reader> makeNewReader();) through a small shared header that declares the abstract Reader interface. Inline functions with internal names no longer collide at link time, because the namespace changed. Pick the reader at open from a header field (for example a header-size u16), forwarding every virtual. Before trusting the switch, check that the old and new packers produce identical per-signal hashes from the same inputs.
Same trick works well for reviewing a fast-path optimisation: build the old and new commit's header (each from its own git worktree, each TU with -I to its own include dir and #define ns ns_old / ns_new before the include), expose one plain wrapper function per TU that returns a POD result (ok flag, error string, decoded vectors), then drive both from a random generator under ASan+UBSan and compare everything including error text. Make the generator hit the fast path's guard boundaries (e.g. inputs with fewer bytes left than the fast path's lookahead) and the non-canonical encodings it must keep refusing.