[bitnom]

The ghost silently fails to appear in two situations. Neither logs anything and neither errors — the request runs, the status indicator shows `suggesting...`, and `renderGhostText` just returns the lines untouched.

1. pi-tui >= 1.4 changed how the cursor cell is encoded

pi-tui 1.4 wraps the drawn cursor in zero-width APC markers and resolves them to inverse video later:

```js
export const CURSOR_MARKER = "\x1b_pi:c\x07";
export const FAKE_CURSOR_START = "\x1b_pi:fc\x07";
export const FAKE_CURSOR_END = "\x1b_pi:/fc\x07";
// resolveFakeCursors() later maps these to \x1b[7m / \x1b[27m
```

The cursor skip assumed an SGR sequence. On an APC marker it consumed one ESC byte, then failed the trailing-padding check and bailed.

2. A framing extension leaves a border at end of line

pi-pretty-tui patches `CustomEditor.prototype.render` — the very class this editor extends — to re-render at `width - 2` and wrap each row:

```js
const innerWidth = width - 2;
const lines = originalEditorRender.call(this, innerWidth);
if (index < bottomBorderIndex) return this.borderColor("│") + fitted + this.borderColor("│");
```

That leaves `│` after the padding, so requiring "spaces to end of line" rejected it. Reproduced by copying that patch verbatim:

```
plain layout -> ghost drawn: true
pretty-tui -> ghost drawn: false (cursor row ends with │)
```

Fix

One matcher skips SGR sequences and APC strings alike:

```js
const DECOR = String.raw`(?:\x1b[[0-9;:]m|\x1b_[^\x07]\x07)`;
```

The padding match now captures whatever chrome follows the blank run, reserves its visible width so the ghost can never overdraw the frame, and re-emits it unchanged. A row with no usable blank run, or with a second cursor marker, still bails — so the original "don't touch a layout you don't understand" guard survives for genuinely unknown content, while a border after real padding is correctly treated as chrome.

Verification

18 tests pass. New cases cover the APC spelling, a styled frame border (SGR-wrapped, as `borderColor` produces), and confirm the untouched-content guard still refuses.

Both failures are invisible at runtime, which is what made them expensive to find; the tests are the guard against a third one.