fix(cursor): z-stacking suppression for occluded cursors (A24)

[re-marked]

Closes #83 (A24).

What

Cursor declarations now follow z-stacking. At finalize time in `ink.tsx`, hit-test the declared screen cell. If the topmost painted element isn't the declared node or a descendant of it, suppress the cursor — something else paints on top.

Symptom this fixes: a background ``'s bar/block cursor renders through any modal-shaped Box on top, even though the modal occludes the cell visually. Per Mark's repro at #83, the cursor visibly bleeds inside the modal frame.

How

• New `cursor-visibility.ts` with `isDescendantOrSelf` + `isCursorVisibleAt(root, declNode, col, row)`. Uses the existing z-aware `hitTest` so the predicate matches what the renderer actually paints.
• One call in `ink.tsx`'s cursor-finalize block: if `!isCursorVisibleAt(...)`, set `target = null` (existing code already handles `null` target as "don't park").

Modal-over-input that DOES contain a focused TextInput: the inner input's `useDeclaredCursor` wins last-write-wins; hit-test at the cursor cell returns the inner input (descendant of the modal), which IS the declared node — so the cursor parks correctly inside the modal. No regression.

Tests

11 new tests in `cursor-visibility.test.ts`:

• `isDescendantOrSelf` (4): self, descendant, unrelated, reverse direction
• `isCursorVisibleAt` (7): declared cell on declarer, on descendant, modal-over-input (the A24 repro), modal off-to-the-side, focused-input-inside-modal, off any rendered element, sibling high-z but not covering

728 tests passing (was 717; +11). typecheck / lint / build clean.

Out of scope (future)

• `` opt-out for non-modal subtrees that want to suppress cursors below them. Mark's original A24 spec listed it as a complement to z-stacking, not a replacement; ship as a follow-up if a real consumer asks.
• Cursor-stacking aware of multiple competing declarations (e.g. "multiple TextInputs in different windows, each declares"). Today it's still last-write-wins for the DECLARATION; the new check just gates whether the latest declaration RENDERS. Solving multi-declaration arbitration cleanly is part of the proper Window primitive (A4: No window primitive — Draggable + Resizable don't compose #56).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4a6527d101

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Don't treat layout-only boxes as cursor occluders

When a higher-z absolute Box is used only as a positioning/layout wrapper (no backgroundColor, opaque, border, or child at this cell), it does not actually write any terminal cells over the focused input, but hitTest still returns that Box because the point is inside its layout rect. This makes the new visibility check hide the TextInput cursor even though the underlying input remains visible; for example, an empty transparent overlay wrapper covering the screen will suppress every cursor below it. The occlusion test needs to account for what painted the cell, not just which layout rect contains it.

Useful? React with 👍 / 👎.