test(console): render the panes

[stormer78]

Eleven persona PRs today, and every bug that reached the live console lived in the one layer nothing could test:

what shipped	what saw it
a ref callback that looped the renderer to a blank page (#179)	you, within a minute
a form that opened empty because a selection was never passed to it (#182)	you
a screen that never came back because its mode was decided once (#178)	reading, just in time
a field asking for an identifier the context did not have (#181)	you

The models under all of them — identity-graph, profile-entries, persona-flow, persona-candidates — passed throughout, because each was correct. Rendering is the only thing that sees the rest.

What this adds

tests/harness/ — module hooks and a DOM so node --test can mount a pane. Nothing changes about how tests elsewhere are run or written.

The hooks do two things that are not obvious:

• resolve a ./thing.js import to thing.tsx. The sources use TypeScript's Bundler resolution, which the bundler honours and Node does not; existing tests sidestepped it by importing ../src/.../thing.ts with the real extension, which a component cannot do because its own imports are written the other way.
• transform JSX with esbuild, not typescript. This repo is on TypeScript 7, whose JS API is the native port's small surface — transpileModule and JsxEmit are simply not on it, and a hook written against the 5.x compiler API fails with an undefined enum that names nothing.

The fake agent answers by task URI, returns the relay's real envelope ({ok, result: {kind: "accepted", …}}) so a pane that mishandles a real response cannot pass, and throws on a task the test did not name — a pane asking something unexpected is exactly what a test should notice, so that is a failure rather than an empty answer.

The tests

Nine, each written against the symptom a person saw rather than the fix, so they keep meaning something if the fix is rewritten. Every negative is paired: "a context with no identifier offers to create one" sits beside "a context with identifiers lists them", because the first alone passes against a form that always offers to create.

Mutation-checked against three of the real bugs:

• restoring the scrape-from-a-ref → 3 tests fail (the loop takes the whole pane down, which is what it did)
• dropping the create offer → a context that publishes no identifier offers to make one fails
• ignoring the selection → selecting a persona aims the context's button at it fails

Two harness details that quietly made a test lie

Both cost me a false pass before they were fixed, and both are written down:

• Click a checkbox; do not assign checked. React reads the click. A hand-set value plus a synthetic change looks like a tick to the test and like nothing to the component — the test passed while proving nothing about the handler.
• Text matching takes the innermost element. Document order returns an ancestor, so clicking a card where the test meant a row inside it selects the wrong thing and then asserts against the wrong screen.

Cost

One devDependency (happy-dom), scoped to the extension package. The extension's test script gains tsc -b, matching packages/core's, because the render tests import through the real package and need its dist.

Verification

npm run lint, npm run build, npm test — 865 across four workspaces (extension 260 → 269). All six ci.yml assertions against the real build.