[JakeGinnivan]

Upstreams dddperth#150, and adds the missing half of the fork/core skill pair.

1. fix(a11y): announce new-tab links and the sign-in error

Built and reviewed in the DDD Perth fork, now moved into core so every fork gets it.

AppLink becomes the single link component. It picks <Link> vs <a> from the shape of to and applies the new-tab treatment itself, so a call site can no longer forget rel="noopener" or the announcement. MdxLink and the two other copies of the is-this-external check are gone. The branch has to live in the component because almost every href here is a runtime value (sponsor.website, action.href), so "is this external?" isn't answerable by reading the JSX.

Two details that are easy to get backwards, both now covered by tests asserting the computed accessible name:

• aria-label replaces element content, so an srOnly span inside a labelled icon link is silently dropped. Those call sites need the hint folded into the label instead. Getting it backwards produces markup that looks right and says nothing.
• The separating space is a text node outside the span. Accessible names concatenate each element's contribution with its own whitespace trimmed, so a space inside the span is dropped and the name comes out as "Register as a Volunteer(opens in a new tab)". An NBSP doesn't survive either.

Axe cannot catch any of this — an unannounced new-tab link is a valid link to every automated rule — which is why the coverage is unit tests rather than an added e2e scan.

auth.login's error box gets role="alert" plus focus-on-render. Both halves are needed: the container isn't in the DOM before the error exists, so there was no live region to observe, and the error arrives in a fresh server-rendered document with focus at the top of the page.

The stub's e2e-content-fixture.mdx gains one external link so the visual baselines actually cover the new ↗. Without it the suite passes while rendering no external text link at all — invisible to the very suite meant to catch it.

2. feat(skills): add /core-push, rename /pull-upstream → /core-pull

Changes move in both directions, but only one direction had a skill. The push was rediscovered from scratch each time — including the part that actually matters, which is what must not travel up.

git subtree split --prefix=core works and fast-forwards cleanly onto ddd-core/main, which makes it a tempting way to publish. It is the wrong tool. Verified on this very change: the split would have

• overwritten all 63 of upstream's stub-branded visual baselines with Perth-branded screenshots,
• restored the fork's deliberate nx.json / vitest.workspace.ts deletions, and
• pushed the fork-shape tsconfig.json / vite.config.ts / project.json overrides that /new-conference applies.

So /core-push is a curated PR, and keeps the split only as a cross-check that nothing was left behind by accident.

It also records the layer mapping the layout implies but never stated: conference/content/pages has a counterpart in conference-stub, while the fork's root CLAUDE.md / ARCHITECTURE.md deliberately do not — core has its own, describing the shared upstream rather than one conference. Copying those up is a mistake that looks like diligence. (This PR's CLAUDE.md section was hand-written for core, not copied.)

Renamed for symmetry. The skill description keeps pull-upstream as a trigger phrase so the old name still invokes it, and scripts/sync-core-skills.mjs already removes skills core has renamed away, so forks drop the stale copy on their next pull.

Verification

Check	Result
vitest run	356/356 pass
nx e2e (a11y)	55 passed, 14 skipped
tsc --noEmit	clean
eslint .	0 errors (2 warnings, both in generated worker-configuration.d.ts)
react-router build	succeeds
pnpm vr	72/72 pass, stable on re-run

Baselines, and a bug found while reviewing them

The stub's fixture gained an external link, so that page is ~120px taller.
Regenerated in the pinned Playwright container (v1.62.1-noble), not on the host.

Reviewing the regenerated images turned up a pre-existing bug unrelated to
this change, now fixed here (64ff40e) since it would otherwise keep
corrupting every future regeneration:

The visual baselines were rotting on a timer. Content pages render an
"Important Dates" card whose DaysLeftPill counts down to the conference date,
so /e2e-content-fixture captured a live "N days left" — a fresh pixel diff
every day, on a committed baseline. It surfaced because the first regeneration
here silently folded in 46 days left -> 44 days left, two days having
passed since the previous run. The changed-file count still matched what this
PR predicted, so it looked correct.

The visual-* projects now seed __devDateOverride (the cookie
date-states.spec.ts already uses). Two traps, both of which fail silently —
the cookie is ignored and the page renders at the live clock:

• The domain must come from baseURL. Under pnpm vr Playwright runs
  inside the container and reaches the host as host.docker.internal, so a
  hardcoded localhost is dropped.
• The value must not need URL-encoding. The provider decodeURIComponents
  it, so an ISO datetime's colons must be percent-encoded to survive the cookie
  header. A date-only value sidesteps this.

Because both failures are silent, visual.spec.ts now asserts the countdown
reads the pinned number before trusting any baseline. Verified by breaking the
domain on purpose: the guard fails while all 63 baselines still pass, because
the pill is far below the 2% maxDiffPixelRatio. Without the guard, that break
is invisible.

VISUAL_DATE and its expected countdown live in e2e/routes.ts beside the
other fork-owned fixtures, so a fork edits one file.

33 baselines now differ from main: the 9 content-page-mdx-template
from this PR's fixture line, plus the countdown on agenda, home, sponsors
and content-page at tablet/desktop (the pill is hidden in the small sidebar at
mobile). Every countdown diff is under 0.32%. A clean re-run afterwards passes
72/72 with no churn, which is the point — they are now stable.

Note for reviewers

A fresh worktree has an empty .wrangler/state/, so nx e2e fails the voting tests with no such table: voting_globals until you run nx d1-migrate-local website. Also note nx test depends on build — app/lib/dates/dev-date-override.test.ts greps the built worker, so a bare vitest run fails those 2 tests until a build exists.

https://claude.ai/code/session_012ZUK9WJ1TGNQzBSNwwrBx1