[Tryanks]

In a long conversation, a drag sometimes stops scrolling. The whole viewport slides like a picture instead, its edge clipped and the background showing, and no new rows come into view. It happens most when the drag starts right after a flick, or when you swipe quickly up and down. Reported on a phone and reproduced on a macOS trackpad.

The cause is in gpui-base 0.8's own wheel handling (ScrollBounce with scroll_wheel(true), the default since #922). It stretches the edge away from both ends in two ways:

• Overscroll::packet gives a release the finger's speed even when nothing is stretched, and Physics::let_go throws the edge from position 0. For about half a second after a flick, the content sits up to 36 px off where it scrolled (with no limit on touch). A gesture that starts in that window grabs the stretch and stretches it further, so the list doesn't move.
• The amount the content used is computed as an f32 difference, (before + dy) - before. At an offset around 8,800 px this can come out a few ten-thousandths of a pixel below dy, and unused != 0. reads that as reaching the edge. Every later packet in the same direction then goes into that stretch.

Logged in the running app (offset −8082.5 of 8902.5, so far from either end): after a flick, a new gesture's +11, +34, +23, +39 px packets left the list offset at −8082.5 while viewport_top drifted from 70 to 80.

page_viewport now passes scroll_wheel(false). The content then scrolls itself, and ScrollBounce stretches only by what's left over when the list is exactly at an end (at_outward_edge && residual.abs() > 0.01). A release lets go only of a stretch that exists. Edge bounce stays. This override is the one page_viewport owns for every page-level area: timeline, sidebar, settings, palette, diff list and panels. Per principle 6, a comment at the site names the gpui-base gap, and the override goes once gpui-base closes it.

Tests:

• a_drag_right_after_a_flick_scrolls_the_timeline (new) flicks six 70 px trackpad packets through the middle of the timeline, one per displayed frame. It then puts the fingers back down and drags 200 px. Without the change the drag scrolls 0 px (left: -420px) and the viewport sits 25 px off. With the change it scrolls 200 px and nothing is displaced. Existing coverage didn't reach this: the edge-bounce test covers only the ends, and the paging test settled frames before every gesture.
• pages_read_far_up_are_released_once_the_reader_stays_at_the_tail gets back its pre-Upgrade to gpui-base 0.8.0 and adopt its scroll, resizable, input and webview defaults #922 driver, which sends one trackpad gesture with no settle_frames() before it. Upgrade to gpui-base 0.8.0 and adopt its scroll, resizable, input and webview defaults #922 added that settling because the test "stalled by starting new gestures against a displaced edge". That stall was this bug: without the change, the restored driver hangs again. With it, the test passes, three runs out of three. A real reader doesn't wait for the edge to settle before the next gesture.

Checks run:

• cargo fmt --all --check and git diff --check.
• cargo clippy --workspace --all-targets --locked -- -D warnings.
• cargo nextest run --workspace --locked.
• cargo check -p tcode-ios --target aarch64-apple-ios-sim --locked and cargo check -p tcode-web --target wasm32-unknown-unknown --locked, both with RUSTFLAGS='-D warnings'.
• Manual: on a macOS trackpad, quick swipes up and down in a long conversation no longer slide the viewport. The app log for that run has 736 trackpad events. Every displaced frame starts at an end of the list: the last frame or two of a real edge bounce springing back.

Not run: the Android check (no NDK on this machine; CI runs it), and a device test on a phone. The sidebar, settings and other page areas get the same change, but I only felt it by hand in the timeline.

• This change alters the wire protocol, so the next release needs a
  PROTOCOL_VERSION bump: a note was added under "Unreleased" above the
  constant in crates/protocol/src/lib.rs (the number itself changes only
  when the release is cut — CONTRIBUTING.md, principle 9).