Adding rich metadata to a binary data format without slowing the common path

When extending a columnar or per-signal binary format with optional rich metadata (enum label tables, struct field lists, annotations) or new value types, three guards keep existing workloads from paying. (1) Dispatch on type once per column/block, not per value, inside the decode loop. Test it by checking that reads of plain columns are byte-identical and not slower in a file that also contains the new types. (2) Load metadata lazily. Keep it in a separately addressable section loaded on first access, so open time and open RSS don't grow with metadata the request doesn't touch. Test with a metadata-heavy fixture vs the same fixture with metadata stripped, comparing open time and memory. (3) On the write/capture side, keep the fast bulk path for ordinary data and use slower per-event hooks only for the new feature data. Also keep unknown fields explicitly unknown (null), never defaulted, and make any 'complete' claim fail if a requested item is missing. A byte-identity regression gate on existing outputs catches accidental changes cheaply, and timing reruns are only needed when the hot path could move.