Fixes #949.

Summary

`Ink#fullStaticOutput` is the running concatenation of everything every `` instance has ever emitted during the lifetime of an Ink renderer. It is initialised to `''` in the constructor and only ever grows — no code path shrinks or resets it. When `shouldClearTerminalForFrame` later fires (overflow, leaving fullscreen, fullscreen unmount), Ink writes `clearTerminal + fullStaticOutput + output`, which means any item emitted by a `` that has since been unmounted/replaced re-materialises above the current viewport even though it no longer exists in the React tree.

Real-world symptom: "cleared history reappears after a long response scrolls the viewport" in TUIs that drive a Static replay via a remount key (Gemini CLI, qwen-code, …). End-to-end repro thread: QwenLM/qwen-code#4083.

Fix

Track `rootNode.previousStaticNode` across commits. When `resetAfterCommit` detects that `rootNode.staticNode` has changed identity (first mount, last unmount, or remount via key), fire a new `onStaticChange` hook on the root node before `onImmediateRender`. Ink installs an `onStaticChange` handler that resets `this.fullStaticOutput = ''`. The new instance (if any) re-emits its items into a fresh accumulator on the immediate render that follows.

```diff
// resetAfterCommit, src/reconciler.ts

• if (rootNode.staticNode !== rootNode.previousStaticNode) {
• rootNode.previousStaticNode = rootNode.staticNode;
• if (typeof rootNode.onStaticChange === 'function') {

•  rootNode.onStaticChange();
  

• }
• }
  ```

```diff
// constructor wiring, src/ink.tsx
this.rootNode.onImmediateRender = this.onRender;

• this.rootNode.onStaticChange = this.handleStaticChange;
  ...
• handleStaticChange = (): void => {
• this.fullStaticOutput = '';
• };
  ```

Scope (no behaviour change for the classic case)

• The Tap-style "one Static mounted for the lifetime of the app, items keep getting appended" use case is untouched: `staticNode` identity never changes → `onStaticChange` never fires → `fullStaticOutput` keeps accumulating exactly as today.
• The rootNode.staticNode dangling reference causes OOM after <Static> unmount #904 / fix: clear staticNode reference on Static unmount #905 unmount path is strengthened: in addition to clearing `rootNode.staticNode` (so the renderer stops walking a freed Yoga tree), Ink now also drops the now-orphan output from `fullStaticOutput`, so a later overflow-driven `clearTerminal` rewrite does not resurrect the unmounted Static's items.
• The original `static output stops accumulating after Static unmounts (rootNode.staticNode dangling reference causes OOM after <Static> unmount #904)` regression test still passes — its length-stability assertion (`outputAfterChurn.length === outputAfterUnmount.length`) holds whether `fullStaticOutput` is left intact at unmount or reset at unmount; this PR doesn't change that invariant.

Test

Adds `fullStaticOutput is reset when unmounts so stale items are not replayed` to `test/components.tsx`. Mounts a `<Static items={['HISTORY-A', 'HISTORY-B']} />` in debug mode (every `stdout.write` is `fullStaticOutput + dynamicOutput`), unmounts it, asserts that the subsequent write does NOT include `HISTORY-A` or `HISTORY-B`. Without the fix the assertion fails; with the fix it passes.

The companion remount-via-key case is intentionally NOT tested here because it additionally depends on the `staticNode`-pointer-clobbered-on-remount bug fixed by #948. Each fix stands alone; this PR is independent of #948 and can be merged in either order. The two fixes compose to give the full "remount produces only new items, no resurrected old items" behaviour.

Verified:

• New test fails on `master` (`76d221c`) — confirms the bug
• New test passes with this patch
• Existing Static tests all pass (`static output`, `skip previous output when rendering new static output`, `static output stops accumulating after Static unmounts (rootNode.staticNode dangling reference causes OOM after <Static> unmount #904)`, `render only new items in static output on final render`, non-interactive mode, concurrent mode)
• Other test failures in the full suite (`cursor moves on space input`, `measure-element`-related, `wrap-ansi doesn't trim leading whitespace`) reproduce on `master` without this patch and are unrelated.

Why a hook on the root node and not inline in the reconciler

The reconciler owns the React → DOM mapping; `fullStaticOutput` is a private Ink-instance writer-side accumulator. A hook keeps the layering intact: the reconciler reports an identity change, Ink decides what to do with it. The same pattern is already used for `onRender`, `onImmediateRender`, and `onComputeLayout`.