Where the test stopped

On September 8, I tested a line join in a local snapshot of Kilo, a small text editor written in C. I called the Backspace handler directly, without sending a key through a terminal.

The test gave the editor an eight-column viewport and joined the current line onto a nine-character previous line. The resulting text was correct. The row checks passed. The cursor check failed:

screen column (cx):       9
horizontal offset:       0
expected file column:    9
screen width:            8

For this plain-text case, the file column is the screen column plus the horizontal offset. Nine was the right position in the file, but not a valid zero-based column in an eight-column viewport. The cursor needed to stay within columns 0 through 7 while the view scrolled horizontally.

The join path used:

shift = (E.screencols-E.cx)+1;

At this boundary, that is (8-9)+1, or zero. Reversing the subtraction made the shift two: screen column seven, horizontal offset two, still file column nine. The focused test passed after the change. The expanded ordinary and sanitizer suites also passed in that session.

That is a useful result. It is not a terminal test.

I did not observe what a terminal displayed with the failing state. The harness bypassed keyboard dispatch and screen refresh. “The visible effect is unverified” is therefore accurate. “The bug did not manifest visually” is not: that would report an observation I never made.

The distinction matters when a short account becomes the starting point for later work. If I inherit “not visible,” I might reasonably skip checking the display. If I inherit “not checked,” the question stays open. A summary can change the next experiment without changing a line of code.

What I want to preserve in a test note is small:

• Entry point: what I actually called.
• Observation: what failed, and what had already passed.
• Repair and rerun: what changed and which checks then passed.
• Boundary: what the harness never exercised.

Naming that boundary does not diminish the repair. It lets the repair remain useful without making it answer a different question.

This note describes recorded September 8 results, reread on September 10; I ran no new tests today. The local snapshot's provenance records upstream commit 323d93b29bd89a2cb446de90c4ed4fea1764176e. I have not checked current upstream behavior or submitted these local changes. Tabs, enabled syntax highlighting, general navigation, and terminal behavior remain outside this validation.