Changing a node's type no longer shows validators a node that never existed

Renaming a node's type in place is a supported repair — patch refuses nodeType outright, so a full update is the only way to fix a node that carries the wrong one. Before this change, that repair quietly corrupted what the update's own validators were looking at.

The update pipeline gives validators the existing node and the proposed one typed alike, so a check written as "the proposed version must not be lower than the existing one" can actually compare them. It did that by deserialising the existing snapshot as the proposed content's type — correct, and stated as such in the code, while the NodeType is the same. On a retype it is not: the proposal's type describes what the node is moving to.

Because the platform's JSON reader skips members it does not recognise, that conversion usually succeeded. A node stored as {"title":"Original","sequence":5} retyped towards a record with Title and Reason came out as ("Original", null): a perfectly well-formed value of the new type, carrying the old title, silently dropping sequence, and inventing a null Reason. Nothing threw and nothing was logged. Validators then compared the proposal against that ghost and answered on it. Where the shapes did not line up the conversion threw instead, the snapshot stayed untyped, and every typed comparison skipped its own check and answered valid — a refusal that never happened.

Now the pipeline types the existing snapshot only when the NodeType is unchanged. On a retype it hands validators the snapshot exactly as it read it, and when that snapshot is untyped it logs a warning naming both NodeTypes, the node, and the fact that a typed comparison on that side will skip. Nothing else about the update changes: the retype still lands, and a validator that wants to refuse one refuses it on the NodeType, which both sides always carry.

There is nothing better the pipeline could offer, and that is the point: by the time it sees the snapshot, the node-stream read has already tried the exact recovery — the mesh-wide content-type registry, keyed on the node's own NodeType — and has an outstanding wait for that type to become known. A snapshot still untyped at this point is untyped because nothing in the process can type it. Saying so is worth more than inventing an answer.

The full mechanism, including the case that made the wrong conversion silent, is in Update Validators See Typed Content.