Claim the props a component declares as its own, and add the working agreement

[pathscale]

A compiled component's own props were also being spread onto its root element, so a prop the layout wires up itself was bound twice.

What happens today

The runtime already has the machinery to prevent this. behaviour names the props a component declares, routedKeys excludes them from passthrough, and the root slot spreads only what is left over. The compiler never populated behaviour, so for every compiled component the list was empty and each of its own props landed on the root as well.

For most props that is invisible, just a stray attribute on a wrapper <div>. For a prop named after a DOM event it is not. Given

onInput?: (value: string) => void;

the layout wires that to an inner <input> and calls it with the string, which is what the signature promises. But the same prop is also on the wrapper, and input bubbles. So the inner event travels up and calls the caller's handler a second time with the raw InputEvent. The bubbled call lands last, so that is the one the caller actually observes.

onChange is declared with a translated signature by eight layouts in @pathscale/ui, plus onInput and onSubmit. Consumers had already started working around it without knowing what they were working around:

honey.id/src/features/admin/components/tables/AdminLogsTable.tsx:161
honey.id/src/features/admin/components/tables/AdminEndUsersTable.tsx:211
honey.id/src/features/admin/components/tables/AdminApplicationsTable.tsx:305
nofilter.io/src/features/studio/pages/StudioSettingsPage.tsx:162
nofilter.io/src/features/admin/pages/AdminBackendSettingsPage.tsx:114

All five are typeof value !== "string" guards on Select's onChange.

The rule

Which props are a component's own is a question its props type already answers, and the two halves of the intersection mean different things:

export type PasswordFieldProps = UIBaseProps & {   // referenced: inherited HTML
  value?: string;                                   // literal: this component's own
  onInput?: (value: string) => void;
};

A referenced member (UIBaseProps, JSX.ButtonHTMLAttributes, IconSlotProps) is inherited surface that belongs on the element. An inline object literal is what this component itself accepts and is responsible for placing. So the literal's keys, and only those, are emitted as behaviour.

That keeps ButtonProps, which extends JSX.ButtonHTMLAttributes, passing onClick and the rest through to the element exactly as before.

A props type that cannot be read locally, because it is imported, emits nothing and keeps today's behaviour. That is the safe direction to fail in: it leaves a prop on the element rather than dropping one.

Blast radius

Checked by regenerating a 98-component library and comparing each emitted behaviour list against what its layout actually reads. Of 661 declared keys, the only ones a layout does not place itself are:

• recipe keys, which presentationOwn claims before behaviourOwn (first claim wins), so naming them again is a no-op;
• children, which layouts read through _stable.children;
• a few read through an aliased binding, such as Grid's as via merge({ as: "div" }, props).

No HTML attribute stops reaching an element.

Tests

Three in the transform crate: declared props are claimed, inherited HTML props are not, and a non-local props type emits nothing. One in the runtime covering the actual defect, that a declared prop stays off the root slot while an undeclared handler still reaches the element.

cargo test 66 pass, cargo clippy clean, cargo fmt clean, runtime suite 149 pass.

---

Also in this PR

The repository working agreement (AGENTS.md plus the CLAUDE.md import shim), previously opened as #12 and folded in here so this repository has one pull request rather than two. It is documentation only and touches nothing the compiler change does.

[pathscale]

Added chore(oxc): 0.2.3, following the convention the previous compiler fix used (fix(lint): validate compound recipe slots together bumped 0.2.1 to 0.2.2 in the same PR). Release tag is therefore solid-layouts-oxc-v0.2.3.

Only solid-layouts-oxc needs a version. The change to solid-layouts is a test, and that package publishes dist only, so nothing about it changes on npm.

Worth calling out for review: this is a patch bump but it does change emitted behaviour. A component's own props stop being spread onto its root element, so anything that was accidentally relying on that placement moves. The audit in the PR description found nothing in a 98-component library that does, but it is the kind of change worth a second opinion on the version.