[kalvinnchau]

Problem

A channel switch renders the whole app shell three times. The third render comes from navigation.complete({ status: "opened" }). The controller then publishes a new snapshot that differs only in status, and the shell re-renders about 470 components. The shell never reads "opened": App.tsx and navigation.ts only check status === "failed".

Change

• shellSnapshot(navigation) in src/app/navigation.ts is the useSyncExternalStore getter for useAppNavigation. When a snapshot differs from the one last shown only by "opening" → "opened", it keeps showing the opening snapshot. Every other change is shown as before, including history updates that land on an opened visit.
• ChannelsPage read navigator.snapshot().status during render and relied on that shell render to notice a settled visit. This mattered only when the relay is in error and retrying ("Connecting…" vs "Can't connect. We'll keep trying."). It now subscribes, with a getter that is true only in that case, so ordinary switches don't re-render it.

No behavior change

• Every in-repo reader of navigation status either subscribes or reads only "failed".
• ChannelsPage.test.tsx no longer re-renders the page by hand after the visit completes. It now fails without the subscription.
• A new navigation.test.tsx case checks that an opened visit keeps the shell's snapshot and that a later history change still reaches the shell. It fails if the getter swallows anything beyond the status change.
• Six partial Navigation test mocks gain subscribe.

A plugin page that reads navigation.snapshot().status during render without subscribing no longer gets a free re-render at "opened". No in-repo page does this.

Evidence

The warm channel-switch probe:

• Chromium, 4× CPU throttle.
• The target channel was already loaded.
• Dense short-message history.
• Pointer modality.

Main 9a7c1d228 was compared against this change. 2 Linux and 2 macOS boxes each ran 8 interleaved rounds, giving 24 samples per variant per box. Medians are below, with Mann–Whitney p ≤ 3e-5 for every cell.

box	script	first rows	LoAF blocking
Linux 1	190 → 168 ms (−11.6%)	243 → 216 ms (−11.0%)	161 → 139 ms (−13.7%)
Linux 2	183 → 158 ms (−13.5%)	235 → 212 ms (−9.8%)	152 → 130 ms (−14.4%)
macOS 1	118 → 108 ms (−7.9%)	151 → 140 ms (−7.3%)	70 → 59 ms (−16.2%)
macOS 2	117 → 106 ms (−8.7%)	151 → 139 ms (−8.1%)	70 → 58 ms (−16.6%)

Local tests run:

• vitest for src/bundled/channels/ (1135 passed), src/app/navigation.test.tsx and the other navigation/App suites, src/features/navigation/controller.test.ts and attachment-panes.test.tsx.
• tsc --noEmit and biome check.

🤖 Generated with Claude Code