feat(protocol): honour viewport.resize end to end

[myrddian]

The last piece of the "stop assuming 80 columns" thread, and the first one that crosses into Java.

viewport.resize has been in the v1 protocol all along — declared on both sides, validated on arrival, and then dropped on the floor (RoutingMessageDispatcher → notImplemented). The client never sent it either. So every session rendered against an assumed 80 columns regardless of the actual canvas.

Per ADR-035 this is v1.x behaviour work, not v2 — the message shape already exists, so nothing on the wire changes.

Client (src/viewport.ts)

• Measures one character cell from a probe rendered in the region's own font, divides the main region by it, reports cols/rows.
• Clamps to the server's validated range (20..500 / 10..500) so a very small or very large window can't trip a protocol error.
• Debounces at 150ms and de-duplicates: a drag-resize fires a torrent of pixel events, but the cell count changes rarely, and every report costs a server-side repaint.
• Re-reports on auth.ok and resume.ok. Sends before the socket opens are dropped (WsClient.send has no queue), and a resumed session may never have learned this client's size.

The arithmetic is split into a pure viewportFromMetrics() so it's testable without a layout engine — happy-dom doesn't do layout, so getBoundingClientRect() returns zeros.

Server

• viewportCols / viewportRows / setViewport on VoidCoreSession, defaulting to 80x24 — a client that never reports behaves exactly as before.
• Screen.onViewportResize defaults to a no-op, deliberately not onEnter. Re-entering resets screen-local state: a resize mid-wizard would throw the user back to step 0 (WizardFormApp.onEnter reassigns stepIndex). A resize is not a navigation event. ScreenApp overrides it to recompose and repaint, so mid-flow state survives a phone rotation.
• ctx.viewportCols() / viewportRows() for screens that want to size content.
• The router records the size before invoking the hook, so a screen composing against ctx sees the new width rather than the previous one — there's a test pinning that ordering.

Verification

• Java: 610 tests, 0 failures (119 skipped — the Docker-dependent integration ones)
• Frontend: 80 passing across 12 files, tsc --noEmit clean, bundle builds
• 6 new Java tests, 9 new TS tests

Not verified: a real browser. No actual resize, rotation, or mobile keyboard has been exercised — the measurement path (measureCell) is the part happy-dom cannot cover.

Follow-up, not in scope

Layout.Flow still defaults to DEFAULT_COLS = 80. Migrating flow-rendered screens onto ctx.viewportCols() is the next change — that's what turns "the server knows your width" into "the server uses it". This PR lands the plumbing and the hook; nothing regresses for screens that ignore them.