fix(tui): prevent cursor marker leaks

[muyiyr]

Summary

Fix cursor-marker leakage across the full-screen selection and terminal rendering paths.

• Treat APC cursor markers as positional metadata instead of persistent styling, so sliceWithWidth() does not replay them into later selection slices.
• Preserve the bottommost valid visible marker as the hardware-cursor position and remove duplicate markers from that cursor line.
• Enforce the output-boundary invariant by stripping all remaining cursor markers from every non-image rendered line, including other visible lines and scrollback lines written during a main-screen full render.

This combines the duplicate-marker cleanup reported in #9257 with the upstream selection-slicing cause identified in #9332. The sliceWithWidth() root cause and Ghostty behavior were identified by @Andy8647 in #9332. The duplicate-marker cleanup direction was also independently proposed by @bluefateludi in #9257.

Closes #9257
Closes #9332

Test Plan

• Confirmed the sliceWithWidth() regression test fails before the fix because later slices contain CURSOR_MARKER.
• Confirmed APC cursor markers are not carried into later slices while persistent SGR styling still is.
• Confirmed the regular-screen duplicate-marker test preserves the first marker's cursor column and emits no marker bytes.
• Confirmed a 2-row main-screen render with markers above the viewport and on multiple visible lines preserves the content, uses the bottom visible marker for cursor positioning, and emits no marker bytes.
• Confirmed the full-screen regression starts with one caret marker, performs a real mouse drag after the caret, renders the selection highlight, and emits no marker bytes.
• npm test --workspace @earendil-works/pi-tui
• tsgo -p packages/tui/tsconfig.build.json --noEmit
• biome check --error-on-warnings on the five changed files

The current GitHub Actions failure occurs in the unrelated packages/ai/src/api/google-shared.ts build on FinishReason.TOO_MANY_TOOL_CALLS; current main fails on the same error. The pi-tui build, complete package test suite, type check, and changed-file formatting checks pass.

This PR was prepared with AI assistance and reviewed by me.

[gaoanze888]

There is still a cursor-marker output leak outside the selected line. extractCursorPosition() scans only the visible viewport and returns immediately after finding the bottommost marked line; the patch removes duplicates only from that one line. Any marker on another visible line, or above the viewport during a full render, remains in lines and is passed to the terminal writer.

I reproduced this at exact head b82f251 with a TuiMainScreen whose 2-row terminal renders three lines: offscreen + marker, visible-a + marker, and visible-b + marker. The bottom marker determines the cursor and is removed, but terminal.writes.join("") still contains CURSOR_MARKER; the focused node:test fails on that assertion. This is the same output-boundary invariant the PR intends to guarantee, not just a synthetic duplicate on the selected line.

Please preserve the first valid visible marker for cursor positioning, but strip CURSOR_MARKER from every non-image rendered line before terminal output, and add coverage for markers on multiple lines / above the visible viewport. (The first commit appears to have had broader cleanup and an offscreen regression before the narrowing follow-up removed them.)

[muyiyr]

@gaoanze888 Addressed in f2e059b80.

Cursor detection is unchanged: extractCursorPosition() still scans the visible viewport bottom-up, so the bottommost valid visible marker determines the hardware cursor position. After extraction, applyLineResets() now strips CURSOR_MARKER from every non-image rendered line before either renderer can write it to the terminal. This covers markers on other visible lines and above the viewport during a main-screen full render.

The new regression uses a 2-row terminal with three marked lines: one above the viewport, one upper visible marker at column 1, and one bottom visible marker at column 9. It verifies that:

• offscreen and other visible content are preserved;
• zero CURSOR_MARKER bytes reach terminal output;
• ESC[10G is emitted, proving the bottom visible marker still determines the cursor column.

I also rechecked the original issue paths: same-line duplicates are stripped while the first marker remains the cursor source (#9257), sliceWithWidth() no longer carries APC into later slices while SGR still carries, and the fullscreen drag-selection path emits no marker bytes (#9332).

Verification:

• node --test test/tab-width.test.ts test/tui-render.test.ts test/tui-alt-screen.test.ts — 97 passed
• npm test --workspace @earendil-works/pi-tui — passed
• tsgo -p packages/tui/tsconfig.build.json --noEmit — passed
• Biome on the five changed files — passed

The current GitHub CI failure is in the unrelated packages/ai/src/api/google-shared.ts build, and current main fails on the same FinishReason.TOO_MANY_TOOL_CALLS error; the TUI build completes before that failure.

This comment is AI-generated by Trae CLI.

[gaoanze888]

Verified the fix at exact head f2e059b. Cursor selection still uses the bottommost valid visible marker, while applyLineResets() now strips markers from every non-image rendered line at the terminal-output boundary. The new 3-line/2-row regression covers both another visible marker and one above the viewport, preserves their content, emits no marker bytes, and still positions the cursor from the bottom marker. The full @earendil-works/pi-tui suite passes locally. The GitHub failure is the unrelated packages/ai/src/api/google-shared.ts exhaustiveness error also tracked by #9570, not this TUI change.