feat(offline): implicit net-state model, two-tier version gate, unified lobby

Land the offline-model redesign (ANDROID_PLAN.md O1-O7): replace the explicit offline toggle with a single detected net-state machine, unify the lobby, and add a two-tier client-version gate. Contour-safe: both version vars empty => dormant, the wire change is additive, so web/pwa/vk/tg behaviour is unchanged unless the gate is deliberately configured.

O1 net-state reducer (test-first). O2 store + wiring (connection/offline shims; +@capacitor/network). O3 remove the offline toggle + migrate the pref. O4 two-tier gate: hard update_required degrades to an offline Update/Play-offline notice (not terminal); soft GATEWAY_RECOMMENDED_CLIENT_VERSION -> X-Update-Recommended nudge. O5 unified lobby (device-local + greyed-from-cache server games; self-set identity; closes G-step-0). O6 create flows (with-friends online/offline segment + offline dict guard). O7 docs.

Telegram/VK stay online-only (contour review): the offline model is channel-gated via offlineCapable() (native + plain web; false in the mini-apps). offlineMode.active is hard-gated on it, so the whole model (blue chrome, unified/greyed lobby, transport kill switch, device-local vs_ai/hotseat create) stays inert in Telegram/VK regardless of the detected net state (closing a version-lock path that could have flipped them offline); and New Game's with-friends hides the online/offline segment there, leaving the remote invite alone. ARCHITECTURE already declared "Telegram/VK are exempt" - this makes the code match.

Deploy/CI: wire the version gate through the deploy (GATEWAY_MIN_CLIENT_VERSION + GATEWAY_RECOMMENDED_CLIENT_VERSION as plain unprefixed vars via compose + ci.yaml + prod-deploy.yaml + write-prod-env.sh + .env.example) so the gate is live when set (test-contour commit-hash version fails open; empty => dormant). Disable the manual android-build CI workflow for now (rename .disabled). Bump the app bundle budget 127->130 KB for the added always-loaded wiring. Fix the UI Docker stage (gateway/Dockerfile): install --ignore-scripts so the Alpine/musl SPA build no longer tries to native-build sharp (a local android:assets tool, unused in the image); esbuild's binary is an optional dep so Vite still builds.

Tests: gateway go green (two-tier gate over a real HTTP handler); docker build of the landing + gateway targets green locally; compose --no-interpolate confirms the gateway env; two new telegram.spec.ts e2e (offline signal keeps the chrome online + quick-match opponent choice visible => server enqueue; with-friends shows no pass-and-play), RED-verified; svelte-check 0/0, vitest 617, e2e 248 (chromium + webkit), build + bundle-size 127.8/130.

docs/ARCHITECTURE.md

@@ -100,15 +100,26 @@ dropped). Horizontal scaling is explicit future work.
   auth operations are unauthenticated and return the minted token. A unary
   operation's domain outcome rides back in `ExecuteResponse.result_code` (HTTP
   200); only edge failures (rate limit, missing session, unknown type, internal)
-  surface as Connect error codes. The client treats a connectivity edge failure as
-  **state, not a per-call toast**: a transport `unavailable` or a `rate_limited` flips a global
-  `online` signal that drives a header **"Connecting…"** spinner and softly disables proactive
-  actions, and the transport **auto-retries with capped exponential backoff** — every op on a
-  rate-limit (the gateway rejected it before processing, so it is safe), but only **read-only**
-  ops on `unavailable` (a mutation is never blindly re-sent, to avoid double-applying one whose
-  response was lost — its button is disabled while offline and the player re-issues it on
-  reconnect). A reachability watcher (a lightweight `profile.get` probe) clears the signal when no
-  other traffic is in flight; the live `Subscribe` stream's drop/recovery feeds the same signal.
+  surface as Connect error codes. The client treats connectivity as **state, not a per-call toast**,
+  via a single **net-state machine** — a pure reducer (`ui/src/lib/netstate.ts`) behind a reactive
+  store (`netstate.svelte.ts`); `connection`/`offline` are thin derived shims over it. It has four
+  states: `online`; `connecting` (a call or probe is failing but still inside the anti-flap window —
+  the header **"Connecting…"** spinner, chrome stays online); `offlineNoNetwork` (sustained loss — blue
+  chrome, a local-only lobby, the transport **kill switch**); and `offlineVersionLocked` (the gateway is
+  reachable but the build is below the hard minimum — the *update* notice, the client-version gate below). **Offline is
+  implicit** — detected from failed calls, a reachability probe and the OS hint (`navigator` /
+  `@capacitor/network`), **never a user toggle** — with **hysteresis** so a brief blip lives entirely in
+  `connecting` (K consecutive probe failures *or* a debounce window trips `offlineNoNetwork`; the first
+  probe/call success heals back, with a toast). The transport **auto-retries with capped exponential
+  backoff** — every op on a rate-limit (the gateway rejected it before processing, so it is safe), but
+  only **read-only** ops on `unavailable` (a mutation is never blindly re-sent, to avoid double-applying
+  one whose response was lost — its button is disabled while offline and re-issued on reconnect). A
+  reachability watcher (a lightweight `profile.get` probe; a session-less native guest reconciles a
+  server guest instead — that reconcile IS the probe) drives recovery, and the live `Subscribe` stream's
+  drop/recovery feeds the same machine. **Telegram/VK are exempt** — always online: they are never fed an
+  offline signal, and `offlineMode.active` is additionally hard-gated on `offlineCapable()` (false in
+  those mini-apps), so nothing — not even a version lock — puts them in offline mode: no blue chrome, no
+  local lobby, no transport kill switch and no device-local create paths.
   **Edge hardening:** every request body on the public listener is capped at
   `GATEWAY_MAX_BODY_BYTES` (default 1 MiB — far above any legitimate payload), both at the HTTP
   layer (`http.MaxBytesReader`) and as the Connect per-message read limit, so an oversized
@@ -140,18 +151,26 @@ dropped). Horizontal scaling is explicit future work.
   (your-turn, opponent-moved, chat, nudge). The gateway bridges them to the
   client's in-app stream while the app is open. Out-of-app delivery uses
   platform-native push via the platform side-service.
-- **Client-version gate** (native long-tail protection): a bundled native install (§13) can be
+- **Client-version gate — two tiers** (native long-tail protection): a bundled native install (§13) can be
   arbitrarily old, so the client stamps its build into an **`X-Client-Version`** request header (the app
   version from `pkg/version` / `__APP_VERSION__`), read by the gateway **before it decodes the FlatBuffers
   payload**. The version rides the **outermost stable layer** — an HTTP header, never the FBS payload —
   because protobuf envelopes and headers are version-tolerant by design while the FBS payload is the layer
-  that breaks. Enforcement is **per-call, not a Hello RPC**: `Execute` and `Subscribe` compare the header to
-  `GATEWAY_MIN_CLIENT_VERSION` at the top (before registry lookup / auth / decode), so the first online call
-  an app makes is already gated. A too-old client gets the domain envelope `result_code = "update_required"`
-  (HTTP 200) on `Execute` and `CodeFailedPrecondition` on `Subscribe`, and makes **zero** successful
-  requests. The gate **fails open** (an absent or unparseable header passes) and is **dormant** until
-  `GATEWAY_MIN_CLIENT_VERSION` is deliberately set (empty ⇒ off, so web behaviour is unchanged). The client
-  raises a terminal *update* overlay — native → the store listing (`VITE_RUSTORE_URL`), web → reload.
+  that breaks. Enforcement is **per-call, not a Hello RPC** (`Execute`/`Subscribe` check it at the top,
+  before registry lookup / auth / decode), so the first online call is already gated. Both tiers **fail
+  open** (an absent or unparseable header passes) and are **dormant** until deliberately configured (both
+  vars empty ⇒ off, web behaviour unchanged).
+  - **Hard (critical) — `GATEWAY_MIN_CLIENT_VERSION`**: `version < min` ⇒ the domain envelope
+    `result_code = "update_required"` (HTTP 200) on `Execute` and `CodeFailedPrecondition` on `Subscribe`;
+    the client makes **zero** successful requests. Its reaction is **graceful, not terminal**: it enters
+    `offlineVersionLocked` (offline mode) and shows a **dismissable notice** — **"Update"** (native → the
+    store listing `VITE_RUSTORE_URL`, web → reload) or **"Play offline"** (dismiss → keep playing local
+    vs_ai / hotseat). The lock is sticky until an actual update on the next launch.
+  - **Soft (recommended) — `GATEWAY_RECOMMENDED_CLIENT_VERSION`**: `min ≤ version < recommended` ⇒ the
+    gateway sets an **additive** `X-Update-Recommended: 1` **response header** on the served `Execute`
+    response (the call still succeeds); a transport interceptor turns it into a **dismissable "update
+    available" nudge** in the lobby — play continues, nothing blocks. `recommended` must be **≥ `min`**
+    (validated at config load); empty ⇒ the soft tier is off.
 - **Frozen wire contract** (so any build, however old, can always recognise "update required"): three
   things are permanent — (1) the protobuf envelope field numbers in `edge.proto` are never renumbered or
   reused; (2) the `update_required` sentinel — the `result_code` string **and** the `FailedPrecondition`
@@ -160,10 +179,12 @@ dropped). Horizontal scaling is explicit future work.
   the FBS payload, which is exactly what the gate guards. **Deploy discipline:** the production rollout that
   ships an incompatible wire change **also** bumps `GATEWAY_MIN_CLIENT_VERSION` to that release, in the same
   rollout (see [`../deploy/README.md`](../deploy/README.md)).
-- **Gate × offline** (UX rule): deliberate offline uses the network kill switch, so the gate never fires
-  while playing offline — an old install can always play local vs_ai / hotseat. The terminal *update*
-  overlay is raised **only on a user-initiated online action**; the silent background guest reconciliation
-  (§3) swallows an `update_required` and stays a local guest rather than interrupting local play.
+- **Gate × offline** (UX rule): offline is the net-state machine's `offlineNoNetwork` / `offlineVersionLocked`
+  (implicit — no toggle) and its kill switch refuses calls, so the hard gate never fires *while already
+  offline* — an old install always plays local vs_ai / hotseat. A hard `update_required` on a
+  **user-initiated online call** degrades to `offlineVersionLocked` + the dismissable notice (above); the
+  silent background guest reconciliation (§3) **swallows** an `update_required` and stays a local guest
+  rather than interrupting local play.
 
 ## 3. Authentication & sessions
 
@@ -1369,13 +1390,18 @@ route (its `directoryIndex` is `index.html`) would otherwise shadow it and serve
 cache-first — the old build's version until a second load. The landing page and the conditional
 polyfill bundle are excluded, and the Connect stream and runtime API POSTs are never precached nor
 intercepted, so the live app is never served stale. This satisfies Chromium's installability requirement (a registered SW, needed
-for install on Android) and powers the opt-in **offline mode** (in progress): a deliberate, device-scoped Settings
-toggle — distinct from the transient gateway-reachability signal — that tints the header blue with
-an *Offline* chip and confines play to on-device `vs_ai` games. The **offline lobby lists only those
-device-local games** (reconstructed by replaying the IndexedDB move journal) and its New-vs-AI entry
-creates one through the in-browser engine — the same game screen then drives it, the robot replying
-locally; online-only affordances (the Stats tab, the random-opponent option) are disabled
-or hidden, and New Game's *with friends* becomes the entry to **local pass-and-play (hotseat)**.
+for install on Android) and powers **offline play**. Offline is **implicit** — the net-state machine
+(§2) detects lost connectivity and tints the header blue with an *Offline* chip; there is **no toggle**
+(the old deliberate Settings switch and its cold-start dialog are gone). The **unified lobby** merges the
+device-local games (active — reconstructed by replaying the IndexedDB move journal) with the last-cached
+**server games shown greyed** (un-openable, a tap toasts *offline*); the Stats tab is disabled and
+invitations are hidden while offline. On offline-capable channels (native / plain web) New Game's *with
+friends* carries an **online/offline segmented control** — online = a friend invite, offline = **local
+pass-and-play (hotseat)** — with the online segment disabled and offline forced when there is no network;
+the online-only Telegram/VK mini-apps skip the segment and show the remote invite alone. A `vs_ai` or
+hotseat create is guarded when
+the chosen variant's dictionary is not available offline. A device-local `vs_ai` game is created through
+the in-browser engine and driven by the same game screen, the robot replying locally.
 A hotseat game is a device-local **2-4 human** game built through the same engine (`hotseat` +
 per-seat/host PIN locks on the record; the seats carry names but no accounts). A **mandatory host
 (referee) PIN** gates the roster at creation and, in-game, the referee overrides — **skip** the
@@ -1403,27 +1429,17 @@ eligible installed PWA (standalone web + confirmed email) **background-preloads*
 — on lobby entry and on a variant-preference change — through the same three-tier loader, retried
 with backoff and honouring the session miss-breaker; the move generator, the loader and the preload
 orchestration stay in lazy chunks. A first-lobby preload failure shows a *poor-connection* notice in
-the ad-banner slot. Flipping the Settings toggle **to** offline runs that same cache-first fetch
-bounded by a ~5 s UI wait (`raceOfflineReady` + the lazy `dict/offlineready`): it enters offline only
-once every enabled variant is ready, otherwise it stays online with a *needs internet* note while the
-fetch finishes in the background (the next flip is then instant). A **cold launch already in offline
-mode** boots from the persisted session and
-profile (the profile is saved on every online adopt/refresh) — `bootstrap` skips the session
-adoption and profile fetch that would otherwise hang with no network, and lands straight in the
-offline lobby; without a cached profile (an install that was never online) it drops the sticky flag
-and boots online instead. Beyond the sticky toggle the app **auto-detects connectivity**: the
-reactive flag carries an *auto* bit, so a self-entered offline (no network) is transient
-(session-only, never persisted) and self-heals to online when the network returns, while a deliberate
-one (the toggle, or the cold-start dialog's *Switch*) persists. At cold start `navigator.onLine ===
-false` enters offline for the session; otherwise a single bounded reachability probe (a `profile.get`,
-~3 s) decides — success boots online, a timeout on an eligible install raises a *No connection* dialog
-(*Switch* = sticky offline, *Wait for network* = boot online with the reachability watcher retrying).
-Mid-session the `window` `online`/`offline` events drive it — `offline` auto-enters, `online`
-re-verifies reachability before returning — backed by a `navigator.onLine` poll because those events
-are unreliable on some platforms (notably iOS PWAs). Offline is a real **transport kill switch**
-(every gateway call is refused, so no traffic leaks); the reachability probe is the one call exempt
-from it (it is the return-to-online mechanism), and the transient reachability watcher is suppressed
-while offline. The gateway registers the `.webmanifest` MIME type
+the ad-banner slot. A **cold launch with no network** boots offline-first from the persisted session +
+profile (saved on every online adopt/refresh) — `bootstrap` skips the session adoption and profile fetch
+that would otherwise hang, and lands straight in the offline lobby; a native launch with no server
+session enters as a device-local guest (§3), and any stale pre-redesign offline-preference key is cleared
+on boot. Connectivity is **detected, not chosen** (the net-state machine, §2): a cold `navigator.onLine
+=== false` or a failed cold reachability probe (a bounded `profile.get`) enters offline, and mid-session
+the `navigator` / `@capacitor/network` `online`/`offline` events plus the probe watcher drive it — with
+**hysteresis** so a brief blip lives in `connecting` and never flips the chrome, and **self-heal** (a
+back-online toast) when the network returns. Offline is a real **transport kill switch** (every gateway
+call is refused, so no traffic leaks); the reachability probe is the one call exempt from it (it is the
+return-to-online mechanism). The gateway registers the `.webmanifest` MIME type
 in-process (the distroless image has no `/etc/mime.types`). Hash-named `/assets/*` are served
 `immutable` (a relaunch is a cache hit, not a re-download); the HTML shells are
 `no-cache` so a new deploy is picked up — both containers apply the same caching. An