Recursive-by-value nodes and string-initialised numeric slices; and why the keyword field-name fix is blocked

[gHashTag]

Two corpus defects that deliberate failing tests exposed, both fixed here, and one emitter fix that is entangled with an open decision and is therefore not.

Recursive by value — 9 fields in 6 files

A node type that contains itself by value has infinite size and cannot be named:

pub const Tree(T) = struct {
    left : "Tree(T)",
    right : "Tree(T)",
};

?T does not help — an optional of a value is still the value's size plus a tag — so ?ListNode and [8]?OctNode are the same defect wearing a question mark. All nine become ?*T, one word and terminating.

Found by following the chain rather than grepping one file: A holding B holding A is the same defect and a single-file scan misses it. Six declarations: Tree(T), List(T), ListNode, KDNode, OctNode, QuadNode.

Measured by running: tri/trees/tree.t27 went from 1 test passing, 1 failing on the infinite-size claim to 2 passing. octree 4 passing, quadtree and kd_tree likewise nameable.

A numeric slice initialised from a string — 4 lines in 3 files

const DEFAULT_KERNEL : []u32 = "[2, 2]";

The string already contains a literal of the right shape, so unquoting is the whole fix. avgpool2d_layer now runs 6 tests.

Neither shows up in zig ast-check: 290 valid and 594 errors are unchanged, because these are semantic, not syntactic. That is the third instrument earning its place.

The keyword field-name fix: correct, and reverted

specs/tools/schema.t27 declares a JSON-Schema record with a field literally named enum. The parser accepted a field name only when the token kind was Ident, so the name was dropped and the element type became the field name — any: void. 23 such fields across 21 specs, 22 of them enum.

Fixing it is one line, with the colon as discriminator so enum ErrorCode { in the braced dialect is not mistaken for a field. It emits @"enum": ?[any] correctly.

And it costs 5 specs. bson, msgpack, async_stream, sort, color went 0 errors → 1, valid 290 → 285.

The reason is the entanglement with #2710: those five contain enum : [Ascending, Descending], the damaged shape that means an enum. Today the field name is dropped, and the leftover bracket list is emitted as Ascending: void, Descending: void — a struct that compiles and satisfies the pinning tests. Preserving the field name emits @"enum": [Ascending,Descending], which is not a type.

So the current passing state of those five depends on a name being thrown away. Both available fixes break them: rewriting the shape to an enum (refuted in #2710 by 27 use sites) and preserving the field name (this). The 23 keyword fields cannot be repaired until enum : [A, B] has an owner's answer.

Reverted with git checkout HEAD -- and rebuilt; 290 / 594 confirmed restored before continuing.