feat(live-activity): honour the streamer's liveActivityPush flag

[RonenMars]

The client half starts a Live Activity locally from session_update frames, independently of whether the server can push to it.
Against a server with liveActivityPush off that produces the worst version of the feature: a Lock Screen card that only updates while you are already looking at the phone, then freezes the moment the app backgrounds and expires silently at iOS's ~8h cap.

Both reconcilers now return early unless the server opted in.
Android's ongoing notification is included — it renders the same feature, so one flag has to govern both.

Why the lookup is synchronous

reconcile() runs on every session_update frame and cannot await an HTTP round-trip per frame, so isLiveActivityEnabled(serverId) answers from a per-server cache and kicks the fetch off in the background.

• Unknown reads as off, matching the streamer's registry default and erring toward not drawing a surface we may have to tear down a moment later.
• One request per server: a burst of frames collapses into a single in-flight lookup.
• A failed lookup is not cached, so one dropped request cannot pin the feature off until the app restarts. The next frame retries.

The one behaviour change worth flagging

A server too old to serve GET /api/config/feature-flags reads as off rather than being grandfathered on.
An old streamer is precisely one where the push half cannot exist, so "on" there would guarantee the degraded local-only path — the exact thing this flag exists to stop.
It also keeps the upgrade direction sane: grandfathering would mean upgrading your streamer removes a visible feature.

Because that is the only path where surfaces stop appearing without anyone choosing it, it logs liveActivity.legacyServer with the serverId.
From the phone the absence looks identical to a server that answered false, and that log line is what separates "upgrade the streamer" from "check the flag".

Server side

Requires RonenMars/threadbase-streamer#545, which adds the flag and gates the server half.
Until that lands, every server 404s the endpoint and reads as off.

Verified state

tsc --noEmit clean, eslint clean on all changed files.
Full unit suite: 1006 passed, 106 suites, 0 failed.

New coverage: 7 cases for the gate module (unknown-then-resolved, explicit off, legacy 404 + its log, per-server independence, request collapsing, retry-after-failure), plus reconcile-level on/off cases for both platforms.
The iOS suite had never exercised reconcile() at all — only the pure helpers — so those two cases are new ground.