cmd/ctrl + scroll now pans the canvas up/down instead of zooming at the
cursor. The catch: a trackpad pinch reports as a wheel with ctrlKey set,
indistinguishable at the event level from a real Ctrl+wheel. Gate on the
physically-held key (cmdRef, tracked via keydown): a real modifier down
pans; a pinch has ctrlKey without any keydown, so it still zooms at the
cursor. Plain wheel continues to zoom at the viewport center.
The webview-forwarded cmd/ctrl+wheel path (canvas-wheel-zoom) stays
cursor-zoom — that channel can't tell a pinch from a held key, and
keeping it zoom preserves pinch-to-zoom while hovering a browser card.
A mouse-wheel notch arrives as deltaY 100 where a trackpad sends ~1-10;
piped through the zoom curve at default sensitivity that's a ~24% jump
per notch, and macOS wheel acceleration stacks them. Clamp the per-event
zoom delta to +/-24 so each notch is a small predictable step. No-op for
trackpads (their deltas are already under the cap), so the continuous
pinch/scroll curve is unchanged.
Plain vertical wheel/two-finger scroll over the canvas now zooms
(anchored at the viewport center, reusing the zoomIn/zoomOut anchor)
instead of panning the y-axis. Horizontal-dominant scroll still pans X,
gated on the dominant axis so a sideways swipe's vertical jitter can't
also zoom. cmd/ctrl+wheel and trackpad pinch still zoom at the cursor.
flushWheel now folds a same-frame pan into the zoom branch, since a
vertical zoom and a horizontal pan can now be accumulated in one RAF
tick; dropping the pan would swallow the gesture.
The Downloads scan rejected any file whose mtime was newer than the scan's
Date.now() snapshot (sub-ms fs timestamps, NTP stepping the clock back), and
the ambiguity test was green only because both fixtures were skipped by that
guard. Allow 60s of future skew, add a win32 end-to-end bind test, and cover
every stamped artifact name (mac arm64/x64 dmg, windows setup exe, linux
AppImages) plus unstamped negatives in the parser test.
Electron resolves ONE install id before spawning the backend (install.json
app_install_id -> python settings installation_id -> fresh uuid) and exports
it as OPENSWARM_INSTALLATION_ID; the backend adopts it when settings has no
installation_id yet. The affiliate app_install_id and the analytics
install_id are now the same value, so affiliate refs join directly to
telemetry with no sign-in required. Drops the frontend app_install_id
forwarding through sign-in flows; the desktop auth router forwards
settings.installation_id as install_id instead.
The optimistic spawn placed the chat with placeCard (which grid-snaps + collision-dodges) then immediately overwrote it with setCardPosition to force the exact viewport-center/beside-card spot. Fold that into a single placeCard({ ..., exact }) so the pre-resolved position is honored directly; the single-browser dock omits exact to keep the collision-dodge cascade.
SirKentut