fix(app): keep the unread session indicator stable during polling

[chphch]

The unread session indicator flickers off and back on during normal polling, with no user interaction.

buildSessionListViewData() bakes each row's hasUnread into the list snapshot, reading it from an optional unreadSessionIds set. applySessions() passes the set, but several other reducers rebuild the snapshot without it — most importantly applyMachines(), which runs on every machine heartbeat (update-machine, machine-activity) and the periodic fetchMachines poll. On each of those rebuilds hasUnread falls back to undefined?.has(id) ?? false, so every row loses its unread state until the next applySessions() restores it. Because machine heartbeats arrive on a timer, the unread indicator (the status dot/text on main) blinks off→on repeatedly while the user just watches the list.

The fix makes unreadSessionIds a required parameter on buildSessionRowData() / buildSessionListViewData() so any reducer that forgets it becomes a compile error, and threads state.unreadSessionIds through the four reducers that dropped it: updateSessionDraft, applyMachines, deleteMachine, deleteSession. (applySessions, markSessionRead, markSessionUnread, and setCurrentViewingSession already passed it.)

Repro: open the sessions list with at least one unread session and an online machine; the unread indicator blinks each time a machine heartbeat arrives (a few times a minute). After the fix it stays stable.

Verification: pnpm typecheck is clean. Making the parameter required is what guarantees completeness here — every call site is now compiler-checked. A running-app capture is awkward for a sub-second timer transient; happy to add one if you'd prefer it over the mechanism walkthrough above.

[chphch]

Rebased onto current main (3c0155c2) — was conflicting. Force-pushed, single commit.

The conflict was stale code on this branch's side: it cleaned up per-session model/effort picks from MMKV in deleteSession, and main has since moved those picks into synced session metadata (#1492), so the MMKV helpers are gone. Dropped those lines; the rest of the commit applied.

Worth noting what main has already picked up in the meantime: two of the five rebuild sites (updateSessionDraft, deleteSession) now pass unreadSessionIds. What this PR still does on top:

• makes the parameter required on buildSessionRowData / buildSessionListViewData, so a reducer that forgets it is a compile error rather than a silent badge wipe;
• threads it through the three that still omit it — applyMachines (the one that fires on every machine heartbeat, so the dominant flicker trigger), deleteMachine, and the draft rebuild.

pnpm typecheck clean, app suite 779/779.

Heads-up on overlap: #1580 (read-position unread sync) also makes this parameter required as part of changing where hasUnread comes from. This PR is the smaller, independent half — if you take it first, #1580 rebases down to just the sync change.