[rcoreilly]

Making a struct got about ten times slower under JavaScriptCore (bun, and Safari) with ef7791d, which stores struct fields as own properties of the instance. The cause is the struct value marker that every generated constructor sets, a symbol-keyed property: on a new object with own field properties, a symbol-keyed store takes a slow path in JavaScriptCore. In base/goscript/gsbench of Cogent Core, LadderStructReturn went from 3,246 to 40,114 ns/op under bun between ec20c61 and ef7791d. V8 (node, Chrome) is not affected. Writing the marker under a literal string name restores the speed; a computed key, even a string one, does not.

The marker is now the string __goStructValue, which the emitter writes by name in the constructor of each struct class. A string key is enumerable, so builtin gains ownKeys, which is Object.keys without the marker, and it replaces Object.keys where the keys of a struct are taken as its fields: reflect.DeepEqual, struct equality, the hashing of struct map keys, print, fmt's struct and object formatting, encoding/json's fallback field list and plain object encoding, hash/maphash, unique, structural type matching and the forwarding of struct views. isEmptyObject skips it.

With this, under bun: LadderStructReturn 40,000 -> 3,166 ns/op, StructCopy 17 -> about 1, StructValueMethod 8.5 -> 3.5, LadderStructValue 11,000 -> 5,200, LadderPointerMethod 13,556 -> 5,330, LadderStructPointer 11,016 -> 5,569, BytesBufferPointer 352 -> 180; the whole gsbench suite is about 10% faster by geometric mean.

The compliance suite, typecheck and vitest pass; the regenerated fixture outputs change only in how the constructors set the marker.

[paralin]

Landed on master as eefc23f, rebased onto the batch #532 to #550, with regenerated fixtures in fca8071. Removed RuntimeHelperStructValueMarker, which the emitter no longer uses, folded the double map in fmt's object formatting into one, and made compareMapKeys (from #548) compare struct keys by ownKeys, so the string marker is not compared as a field.