ADR-0058: Stable transcript reflow and scroll ownership
Status: Accepted · 2026-09-07 · Refines ADRs 0053, 0056 and 0057.
Context
The user reported chat content disappearing during font resizing and scrolling. A native regression exposed a stack-overflow crash in observable font setters that assigned to themselves unconditionally. The transcript also fed estimated lazy-stack heights into immediate scroll commands and animated row placement.
Decision
Use a fully measured VStack containing at most 40 messages, with the existing bounded Markdown cache. Native lazy-stack height estimates remained unstable under large font/width changes, and the native list alternative failed the selection/reflow regressions. A bounded measured window trades implicit whole-history scrolling for predictable geometry and bounded rich-view work.
Earlier and Later controls move by 20 messages with an overlapping window, preserving the current edge as a message-ID scroll target. Latest and a new send return to the newest window. All messages remain in the model and database; model context and generation ownership do not use this display range. Loading/evicting model history remains separate work.
Each message contributes one stable, vertically self-sizing child. Row insertion fades and offsets are removed. Native scroll-position binding preserves reading identity; native size-change anchors handle font, width and prepared-content reflow. Estimated content heights no longer drive explicit scrolling.
A bottom sentinel tracks whether the reader has reached the end. Reader gestures suspend following through deceleration; ending at the bottom resumes it. New sends resume following. Stream-driven commands retain their 15 Hz upper bound and are always deferred out of the current layout frame. Disappearance and reader gestures cancel pending commands.
Font setters assign a normalized value only when it differs, then persist the canonical value. This bounds observable setter reentry while preserving existing preference keys and valid ranges.
Consequences and validation
No storage, network authority, dependency or cache policy changes. Visibility observations stay outside observable view state; resuming following is deferred to avoid layout feedback. Long-history paging and large-code rendering remain PERF-01 and PERF-04.
Native regressions cover long/short chat selection, 80 messages across five width/font combinations and five native wheel movements, and scroll-wheel events followed by streaming and reflow. Font tests cover repeated, fractional, out-of-range and non-finite assignments and persistence. Visual inspection uses the complete hosting view, not its transparent clipping layer.