Motivation

`packages/memtomem/src/memtomem/web/static/` ships ~15 hand-rolled JS
modules (`app.js`, `sources-memory-dirs.js`, `settings-*.js`, `i18n.js`,
etc.) carrying real branching logic — view rendering, sub-tab routing,
filter/sort, orphan grouping, locale resolution. Today the only
automated guards on this code are:

• `test_i18n.py` — locale parity (key set + placeholder shape).
• `test_no_template_literal_toasts` — i18n-coverage AST-grep guard.

That's it. Every behavioural change to the UI relies on manual
verification or Playwright sessions kicked off ad-hoc. Several recent
PRs (#587, #595, #639) shipped UI-rendering bugs whose minimal
repro would have been a 10-line jsdom assertion. The pattern is
explicit in `feedback_claim_test_parity.md`: behaviour claims need
fail-if-false tests on CI, and the static modules currently violate
that.

What this issue is asking for

Stand up a minimal JS unit-test runner that lives next to the Python
test suite and runs in the same CI lane. Constraints:

• No build step. The static modules are served verbatim by FastAPI;
  any test layer that requires bundling drifts from production.
• jsdom or happy-dom. The modules touch `document`, `localStorage`,
  `fetch`. A real browser via Playwright is overkill for unit-level
  rendering assertions.
• Globals friendly. `app.js` exposes `STATE`, `api()`, `t()`, etc.
  on the global scope; the runner needs to either ingest the module
  as-is (script tag style) or accept a thin shim for those globals.

First three test candidates (concrete, not speculative)

1. `_renderMemorySourceTree` with orphan rows — assert that a
   `/api/sources` mock returning `memory_dir=null` produces an
   `.source-vendor-orphan` block under the User vendor, and that
   the user sub-tab badge counts the orphans. Direct regression
   guard for fix(web): render orphan-indexed sources in Sources tab #639.
2. `_renderMemorySourceTree` empty user vendor — assert that
   the empty-state placeholder renders when no dirs and no orphans
   are present, but is replaced by the orphan block when orphans
   exist alone. (Edge case from the fix(web): render orphan-indexed sources in Sources tab #639 review.)
3. `I18N.applyDOM()` with `data-i18n-placeholder` — assert that
   placeholder attributes refresh on language change without the
   static-key clobber bug from feat(web): namespace cluster — preview echo + honest placeholder + helper text (#581) #595 (`feedback_data_i18n_placeholder_clobber.md`).

Suggested shape

```
packages/memtomem/tests-js/
├── package.json # devDeps: vitest, jsdom (or happy-dom)
├── setup.ts # configure jsdom, expose globals
├── _renderMemorySourceTree.test.ts
├── i18n_apply_dom.test.ts
└── ...
```

CI: add a step before the Python pytest run that does
`npm ci && npm test` inside `tests-js/`. Failure fails the same
build that pytest does.

Out of scope

• Full DOM E2E (that's Playwright via `reference_playwright_mcp_web_verification.md`)
• Visual regression
• Changing the no-build serving model