Constant REGISTRY_CRATES 

const REGISTRY_CRATES: [&str; 5];

The crates whose doc comments may not cite the internal corpus, and the crates whose doc comments still do.

M7.D7’s sweep took the internal docs/design/… citations out of the five crates that reach a package registry — the four published to crates.io plus atune_py, whose docstrings become the PyPI page — and left 118 lines in atune_bench, atune_oniro and atune_gui alone on a stated premise: those crates are publish = false and “never reach docs.rs”. Nothing stood guard over the result, so the invariant held only until somebody edited a doc comment. This is that guard, and it is the half that can be enforced today: the five must stay at zero, which they are.

The premise is no longer true, and the other half is §1.3.157. D30 gives every crate a published reference at /atune/<version>/api/rust/ because it never reaches docs.rs, so the same fact that excused those crates is what publishes them. Measured against the real cargo doc --workspace tree: 54 pages carry a docs/design/… path — atune_bench 21, atune_dev 23, atune_oniro 9, atune_gui 1 — which is the §15 violation the sweep exists to prevent, on the pages D30 created. atune_dev is not even in the plan’s count; it postdates it. Repairing them is prose work with a per-citation judgement (§9’s rule table: a user-relevant fact becomes a guide link, a build note moves to a plain //), so the counts are printed on every green run rather than gated: a number that shows up on every run is a debt somebody can watch shrink.