[stupart]

§3: "A hairline appears under the header only once the transcript scrolls or a pane opens." It was drawn unconditionally, so a header with nothing above it still carried a rule — a line separating nothing from nothing.

The signal was already being measured

publishPosition() computes fromTop to decide when to ask for the previous page (onReachTop). The hairline asks a different question of the same number — has anything gone under the header yet — so this adds one call beside the existing one rather than measuring twice.

It is deliberately not onUserScroll. That reports isAtBottom, which is a different question: a long transcript resting at its top is not at the bottom and has still scrolled nothing. Using it would have drawn the rule immediately on every session with a scrollable history.

The guarded expression onUserScroll(document.height <= visible.height || distance <= 8) is untouched — it is what keeps a streaming row followed.

Both of §3's reasons

The conversation page waits for the scroll. The deliverable and side-by-side pages keep their rule unconditionally, because there a pane is open — §3's other clause — and a guard asserts that arm never gains the condition.

Mutation battery, including one that got away

mutation	result
the hairline is drawn unconditionally again	1 fail
the pane stops listening for the scroll	1 fail
any scrollable transcript counts as scrolled, even at rest	1 fail
it asks about the BOTTOM instead of the top	1 fail
the deliverable page's rule also waits for a scroll	1 fail
the measurement is computed a second time	survived at first

The last one is the useful one. My guard counted let fromTop = documentView.isFlipped, and the mutation duplicated the calculation as fromTopCopy — a different name, so the regex correctly didn't match and the guard passed. That is also the realistic drift: someone adds a second top-distance under a new name rather than reusing this one. The guard now pins the expression (documentView.isFlipped ? visible.minY - document.minY) instead of the variable name, and the same mutation fails it.

Gate

bun test 2300/0 exit 0 (2299 on main plus this guard) · tsc --noEmit exit 0 · swift test exit 0 · Mac xcodebuild BUILD SUCCEEDED · iOS xcodebuild BUILD SUCCEEDED — the Mac build matters here: a new closure is threaded through an NSViewRepresentable's Coordinator, and CI builds neither app.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SfEX6bfwCGXtThk1JRNmsE