Close the two remaining shadowing holes (§13.25, §13.18)

[kaidesu]

Overview

The two items left over from the "does this resolve all shadow/scoping issues?" sweep after #179. One is a new callout that report never found; the other closes a residual #179 recorded rather than fixed.

§13.25 — a library global could not be shadowed by anything

evaluateIdentifier consulted the library registries before the environment, so §9.1's two global names beat every binding in the language — a parameter, a loop variable, a name bound in a block, an import ... as. The binding was made and then never readable.

function render(type) { return type }
render("warning")               // before: library function {type}
                                // after:  "warning"

class Cmd { constructor(type) { this.type = type } }
new Cmd("edit").type            // before: library function {type} — written into the instance
                                // after:  "edit"

Nothing raised, which makes it worse than the shadowing §13.22 described: there the failure was at least a fault at some later call. Here a string parameter silently is a function object and gets stored as one — the silent wrong answer §3 sets out to make impossible. And type is an ordinary word for an ordinary parameter: a command type, an event type, a token type.

§9.1 grants these two global standing "the same standing print/type-checking primitives have in other scripting languages" — where a local shadows them without ceremony (Python's type is a builtin any local shadows). The behaviour contradicted the rationale offered for it.

Fix: the two lookups are swapped — environment first, registries as the fallback when nothing is bound. Shadowing becomes the ordinary rule; a global is what a name means where nothing else claims it. The optimizer's LibraryBinding marking is unchanged and still correct: it now saves an undefined name two map lookups, rather than saving every ordinary variable two, which the new order already does.

Cost: none measurable. Allocation is identical on evaluator/benchmark_test.go and wall time is within run-to-run noise — a global now walks the scope chain before reaching the registry, which is the walk every other name already pays.

§13.18 — the cross-body collision

#179's note recorded this as a residual with the shape of the fix; this closes it. The in-body check couldn't see a field here colliding with a method that arrived from a superclass or a used trait:

class P { label() { return "method" } }
class C extends P { label = 7 }
// before: c.label → 7, c.label() → "method"   (the §13.18 duality, silently)
// after:  syntax error: `label` is declared as a field here and as a method on `P`

checkInheritedCollisions runs once the class body is complete — the earliest either declaration path could see the other, since use may sit anywhere in the body. Both directions are caught (a field here against an inherited method, and a method here against an inherited field), through a superclass, a grandparent, or a trait.

Only field-against-method counts. A name arriving twice as two fields, or twice as two methods, is overriding — the point of extends and use — and five cases pin that it stays legal.

The fault is reported at the declaration in this body and names where the other one came from, that being the half a reader can't see from here. object.Field gained the token it was declared at; declaredMethods reads method names and tokens back off the class body.

Verification

• evaluator/globals_test.go: seven binding forms shadowing a global, an aliased import doing the same, the globals still resolving unshadowed (including inside a function and a method), and a shadow not escaping its scope.
• evaluator/class_members_test.go: five cross-body collisions and the five overriding shapes that must stay legal, alongside the existing in-body cases.
• All 41 programs in examples/ unchanged (mud.gs excepted as before — a non-terminating interactive loop, byte-identical for 2.9 MB, differing only in frame count inside the timeout).
• Studio's 132-case suite passes.
• go build ./..., go vet ./..., gofmt -l ., go test ./... clean.

Note on §15

§13.25 came from sweeping the scoping surface after §14 decision 12, not from the Chisel/Studio report — and it outranked most of what was still open. §15 now records that, because ranking a report is worth doing and is not the same as knowing what is wrong.

With this, §15's next item is #5 (§13.16) — the export surface, and the only ranked item left that isn't small. It needs §14 decision 10 settled before code.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EGsNMRiKwWmvoizxD619zA

---

Generated by Claude Code