fix(idp): login error survives the redirect — ride in lastSubmission — closes #514 - #554

…514)

The failed-login path set `interaction.lastError`, saved, and
redirected back to the form. Every piece of the wiring looked right —
views.js renders the error div, the GET re-render passes the value,
the setters fire — but the error never appeared. Root cause:
oidc-provider's Interaction.save() persists ONLY the fields in the
model's IN_PAYLOAD list (lib/models/interaction.js), and `lastError`
isn't one of them. The property survived in memory, was silently
dropped on save, and the redirected GET re-rendered a pristine form.
Users retried blind; the issue author lost hours to an imagined OIDC
regression (#513) because the IdP refused to surface a wrong-password
error.

Fix: the message rides in `lastSubmission` — the IN_PAYLOAD slot
oidc-provider designates for exactly this form re-render state. All
four sites updated (two setters, the GET reader, and the
switch-account clear), with the IN_PAYLOAD gotcha documented at the
first setter so the next foreign-property bug doesn't take hours to
spot.

Verified live before and after: the issue's repro (wrong password on
a single-user IdP) now renders
`<div class="error">Invalid username or password</div>` on the
re-rendered form.

Regression test drives the real OIDC interaction flow over HTTP
(dynamic client registration → /idp/auth → interaction redirect →
POST bad credentials → follow redirect) with manual cookie threading
and asserts the error div. 948/948 passing.

Closes #514.

…-cookie failure

Copilot: Headers.getSetCookie() landed in Node 18.15/19.7, but
engines still declares >=18.0.0 (bump deferred — #541). The previous
optional-chaining fallback would collect zero cookies on older Node
and fail the test at a confusing distance.

Fix: detect support up front and skip with a self-documenting reason.
Skipping costs nothing on those runtimes — the IdP itself cannot run
on Node 18 at all (oidc-provider uses Array#toReversed and
crypto.hash, #523), so every IdP test is already broken there; this
one now says WHY instead of failing mysteriously. The jar call is now
direct (no optional chaining) since the guard owns the decision.