Fixes #947.

Summary

`reconciler.removeChildFromContainer` / `removeChild` unconditionally cleared `rootNode.staticNode` whenever the removed node was marked `internal_static`. React processes insertions before deletions inside a single commit, so when a `` is remounted via a `key` prop change:

1. `createInstance` fires for the new internal_static node and registers `rootNode.staticNode = newNode` (also marking `isStaticDirty = true`).
2. `appendChildToContainer` attaches the new node.
3. `removeChildFromContainer` fires for the old internal_static node, sees `removeNode.internal_static === true`, and zeroes out `currentRootNode.staticNode` — clobbering the pointer the new node just registered.

After that, `renderer.ts` sees `node.staticNode === undefined`, emits no static output, and the new ``'s items are silently dropped from stdout.

Fix

Clear `staticNode` only when the node being removed is the one the root currently points to. Identical guard added to both `removeChildFromContainer` (root-level remount) and `removeChild` (nested remount).

```diff

• if (removeNode.internal_static && currentRootNode) {

• if (
• removeNode.internal_static &&
• currentRootNode &&
• currentRootNode.staticNode === removeNode
• ) {
  currentRootNode.staticNode = undefined;
  }
  ```

The existing #904 / #905 path (pure unmount → `staticNode` should be cleared) is unaffected: in that case `currentRootNode.staticNode` still equals `removeNode` at removal time, so the guard passes through.

Test

Adds `remounting via key change emits the new items` to `test/components.tsx`. Asserts that after a key-driven remount of `` with a disjoint items array, the new items reach stdout.

Verified:

• New test fails on `master` (`1510851`) — confirms the regression
• New test passes with this patch
• Existing Static tests all pass — including the rootNode.staticNode dangling reference causes OOM after <Static> unmount #904 `stops accumulating after Static unmounts` regression (the pure-unmount path is preserved)
• Full `ava` suite shows no new failures vs. `master` baseline (two pre-existing flaky failures on `cursor` and `wrap-ansi doesn't trim leading whitespace` are unrelated and reproduce on `master` without this patch)

Real-world impact

Both Gemini CLI and qwen-code use a `` pattern to force a full replay of the chat history after `/clear`, model switch, Ctrl+O, etc. After upgrading to ink 7, the entire history pane stops rendering immediately after the first such trigger. End-to-end repro on qwen-code: QwenLM/qwen-code#4083 (currently has to revert ink 7 → 6 as a stop-gap).