Stabilize main: consolidate WebKit flake and bottom-follow fixes - #1039

Co-authored-by: Mordecai <f314b6033b6e13df94c8e9e1187d3666b6b9ce54e66ee986e31cdaa909e8dcbc@buzz.block.builderlab.xyz>
Signed-off-by: Wes <[email protected]>
(cherry picked from commit c4dba54)

The "Live retry recovers an empty paused roster" journey asserts that Retry
clicks inside the roster quota cooldown send nothing, but it installed a
page clock that keeps running in real time. The app gates the roster read
on performance.now() against a 3s deadline it arms when it consumes the
refusal, so a slow WebKit runner reached the second "ignored" click after
the cooldown had really expired and correctly sent a second roster read.

Hold the first roster read at the page boundary until navigation (which
needs a running clock) is done, pause the page clock, then release it so
the app arms its deadline in controlled time. Only crossCooldown's runFor
now passes that deadline.

Co-authored-by: Wes <[email protected]>
Signed-off-by: Wes <[email protected]>
Co-authored-by: Princess Donut <5d97ac8c272fa949af56c71e586146b1706e5c4c3daa125db5b4625e12e0686c@buzz.block.builderlab.xyz>
(cherry picked from commit ed128d8)

Thread activity in this test is lit only by typing, which stays fresh
for 8s. The detour through the activity panel and back to the thread
can take longer than that on slow WebKit runners, so the thread trigger
was gone by the time the test focused it. Refresh typing before using
the trigger, as each later section of the test already does.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>
(cherry picked from commit e9bb156)

"rejects a full attachment draft before opening the microphone" failed on
WebKit with `"page.evaluate: Test ended." while running route callback`.
WebKit defers the voice card's media reads until the first click and keeps
issuing them, and the shared media route reads its bytes from the page with
page.evaluate. This test ends about 200ms after its only click, so the
context could close under a media read still in flight.

Unroute in afterEach with `ignoreErrors`: those reads serve no assertion
once the test body has finished.

Co-authored-by: Wes <[email protected]>
Signed-off-by: Wes <[email protected]>
Co-authored-by: Princess Donut <5d97ac8c272fa949af56c71e586146b1706e5c4c3daa125db5b4625e12e0686c@buzz.block.builderlab.xyz>
(cherry picked from commit 2607d39)

Co-authored-by: Mordecai <f314b6033b6e13df94c8e9e1187d3666b6b9ce54e66ee986e31cdaa909e8dcbc@buzz.block.builderlab.xyz>
Signed-off-by: Wes <[email protected]>
(cherry picked from commit 307debe)

A helper change makes the selector repeat every journey in the file. For
layout.spec.mjs that is 18 journeys x 10, which ran past the 15-minute job
timeout on WebKit (#1022: 148/170 passed, 0 failed, then cancelled).

The selector now emits one shard per five journeys (at most eight), and the
lane spreads the repetitions with --shard. Small selections keep one runner.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>
(cherry picked from commit 6a85f66)

…not a sync error

Leaving a channel while its catch-up save is queued invalidates the save.
That rejection set the global read-state health to error, so Inbox showed
"Reading observation expired" with Retry inbox, and rows shifted 56px. CI
caught it as an inbox-refresh-layout flake on slow WebKit runners.

The caller still receives the rejection; explicit Inbox actions keep their
own alert.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>
(cherry picked from commit 63f2a1a)

Treat position samples taken from Virtua list-height commits as layout-owned so an upward virtualizer correction cannot masquerade as reader input. Keep explicit upward gestures authoritative, and cover the panel-resize ordering at both the component and browser boundaries.

Co-authored-by: Mongo <81cabd2ca1792494372ed410283c00770c0cf35afca052e940fa522883212b45@buzz.block.builderlab.xyz>

Signed-off-by: Wes <[email protected]>
(cherry picked from commit 85ea95ec179d77bb7ec48c8d2a06ef461370c894)

… bottom

Wes's #1022 P1 case. It fails 120px from the bottom on main in both
engines and passes with the ChannelTimeline bottom-follow fix.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>

…fore its restore frame

#1039's CI still failed the panel-resizing journey 121px from bottom on
WebKit: after a list-height commit, the browser delivered Virtua's 1px
offset correction as a scroll event before the restore frame, and
onScroll read it as reader movement. Treat scrolls in that window as
layout-owned until the restore frame or the next reader input.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>

…to the first slot

The removed photo tile leaves the accessibility tree when its exit
starts but keeps its slot until the exit ends; the clip then slides
left. A click straddling that slide pressed the button and released
beside it. A forced 300ms press failed 6/6 before and passed 6/6 after.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>

…enu journey

"row menu moves and removes a channel optimistically…" failed on WebKit at
openSort: the Sort menu item never appeared after clicking "More actions for
Channels". The click came right after a reload, while the sidebar still
showed "Loading Me placement…". Base UI opens the menu on mousedown and a
document mouseup outside the trigger and popup cancels it; WebKit took
~500ms between press and release, the placement status left during it, the
sidebar shifted up, and the release landed off the trigger.

Open the menu from the keyboard, like openMove already does, and wait for
Base UI to highlight Sort before focusing it. Pointer opening of the section
menu stays covered by navigation-sorting.

Co-authored-by: Wes <[email protected]>
Signed-off-by: Wes <[email protected]>
Co-authored-by: Princess Donut <5d97ac8c272fa949af56c71e586146b1706e5c4c3daa125db5b4625e12e0686c@buzz.block.builderlab.xyz>
(cherry picked from commit fef8a23)

…imeline positions it

On loaded Linux WebKit the timeline's own end alignment can land short
of the DOM bottom while rows re-measure, and onScroll read that
programmatic upward scroll as the reader leaving the bottom. Keep the
timeline-owned position until the next reader input; a held pointer
(scrollbar drag) still lets geometry infer upward movement.

Instrumented in the pinned Linux Playwright image at 1.5 CPUs, the
panel-resizing journey failed 5-7 of 30 before this change and 30/30
passed after.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>

…text

#828's SVG <text> numeral made every interaction near day dividers
about 1.5x slower on Linux WebKit; the crowded nested-replies journey
doubled in CI and pushed WebKit shard 4/8 past its 15 minute limit on
main. Keep the SVG tile and lay an HTML numeral over it. Before/after
WebKit screenshots of the divider match.

Co-authored-by: Carl <32a2e2c9d428ee08902cab75d956da2c1d235a22d4766b0dd4138bf6e2e5db1d@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Wes <[email protected]>