D001 - A fourth repository, holding formats and nothing that knows what a consumer is for

Three projects need these formats and each had grown its own copy. The cost was not hypothetical: a container builder shipped emitting the previous generation's magic, because the current one was recorded in a sibling project's decision log and nothing connected the two. The file would have been rejected by the only machine it was built for, on its first four bytes.

The duplication argument alone would have been arguable - a reader and a writer are genuinely different code, and merging them buys a coordination tax. What settles it is that the duplicated part is the facts, not the code: the header layout, the field offsets, the flag bits and the magic values each existed more than once, and one copy was wrong.

The second reason is contention. A repository is a unit of concurrent work. While these formats lived inside a probe's tooling directory, work on a package writer and work on the probe were the same working copy, so they serialised - one session's refactor blocked the other's feature for an afternoon. Splitting removes a duplication and a queue.

Status: decided.