Stage 6: gateway edge (Connect/FlatBuffers over h2c, platform/email/guest auth, sessions, rate-limit, admin passthrough, live push bridge)

New public ingress and the first network edge. Framework + a vertical slice of
operations end-to-end; remaining ops reuse the same transcode pattern in Stage 7.

Contracts (new module scrabble/pkg):
- push.proto (backend->gateway gRPC server-stream) + scrabble.fbs (FlatBuffers
  edge payloads), committed generated Go; buf/flatc Makefiles (dev-time codegen).

Backend:
- REST handlers on the /api/v1 groups: internal session endpoints
  (telegram/guest/email login -> mint, resolve, revoke) and the user slice
  (profile, submit_play, state, lobby enqueue/poll, chat).
- internal/notify in-process Publisher hub + internal/pushgrpc gRPC server
  (BACKEND_GRPC_ADDR) streaming your_turn/opponent_moved/chat/nudge/match_found;
  emission in game.commit, social, matchmaker.
- migration 00005 accounts.is_guest; guests are durable rows excluded from stats;
  ProvisionGuest; email-as-login (RequestLoginCode/LoginWithCode).

Gateway (new module scrabble/gateway):
- Connect Gateway service over h2c (Execute + Subscribe), FlatBuffers<->JSON
  transcode registry, Telegram initData HMAC validator (seam), session cache,
  token-bucket rate limiter (3 classes), push fan-out hub, backend REST + push
  gRPC client, admin Basic-Auth reverse proxy.

go.work: use ./pkg, ./gateway + replace scrabble/pkg. CI: gateway/**, pkg/**
path filters; unit build/vet/test span all three modules. Docs (PLAN,
ARCHITECTURE, FUNCTIONAL+ru, TESTING, READMEs) updated; gateway/pkg unit tests +
guest/email-login integration tests.

backend/internal/notify/notify.go

@@ -0,0 +1,110 @@
+// Package notify is the backend's in-process live-event seam. Domain services
+// publish Intents after a successful commit; the gRPC push server (internal
+// /pushgrpc) subscribes to the hub and streams them to the gateway, which fans
+// them out to clients (docs/ARCHITECTURE.md §10). Event payloads are
+// FlatBuffers-encoded by the typed constructors in events.go, so the domain
+// services stay free of the wire schema and only depend on this package.
+//
+// Publishing is best-effort and non-blocking: a live event is a convenience, not
+// a correctness requirement, so a slow or absent subscriber never blocks a game
+// transition. The default Publisher is Nop, which keeps every domain service (and
+// its tests) runnable without a live channel.
+package notify
+
+import (
+	"sync"
+
+	"github.com/google/uuid"
+)
+
+// Notification kinds — the catalog in docs/ARCHITECTURE.md §10.
+const (
+	KindYourTurn      = "your_turn"
+	KindOpponentMoved = "opponent_moved"
+	KindChatMessage   = "chat_message"
+	KindNudge         = "nudge"
+	KindMatchFound    = "match_found"
+)
+
+// Intent is one live event destined for a single user. Payload is the
+// FlatBuffers-encoded body (a scrabblefb.* table) that the gateway forwards
+// verbatim to the client; EventID is a correlation id carried through unchanged.
+type Intent struct {
+	UserID  uuid.UUID
+	Kind    string
+	Payload []byte
+	EventID string
+}
+
+// Publisher accepts live-event intents. Implementations must be safe for
+// concurrent use and must not block the caller.
+type Publisher interface {
+	Publish(intents ...Intent)
+}
+
+// Nop is the default Publisher: it discards every intent.
+type Nop struct{}
+
+// Publish discards the intents.
+func (Nop) Publish(...Intent) {}
+
+// Hub is the in-process fan-in/fan-out between the domain publishers and the
+// push subscribers (the gRPC stream). It is safe for concurrent use.
+type Hub struct {
+	mu      sync.Mutex
+	subs    map[int]chan Intent
+	nextID  int
+	bufSize int
+}
+
+// defaultBuffer is the per-subscriber queue depth used when NewHub is given a
+// non-positive size.
+const defaultBuffer = 256
+
+// NewHub returns a Hub whose per-subscriber buffer holds bufSize intents before
+// dropping (a slow subscriber never blocks a publisher).
+func NewHub(bufSize int) *Hub {
+	if bufSize <= 0 {
+		bufSize = defaultBuffer
+	}
+	return &Hub{subs: make(map[int]chan Intent), bufSize: bufSize}
+}
+
+// Publish delivers each intent to every current subscriber, dropping it for any
+// subscriber whose buffer is full (best-effort live delivery).
+func (h *Hub) Publish(intents ...Intent) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	for _, in := range intents {
+		for _, ch := range h.subs {
+			select {
+			case ch <- in:
+			default:
+			}
+		}
+	}
+}
+
+// Subscribe registers a new subscriber and returns its intent channel and an
+// unsubscribe func that closes the channel. The caller reads the channel until
+// it is closed or its own context ends, then calls unsubscribe.
+func (h *Hub) Subscribe() (<-chan Intent, func()) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	id := h.nextID
+	h.nextID++
+	ch := make(chan Intent, h.bufSize)
+	h.subs[id] = ch
+	return ch, func() { h.unsubscribe(id) }
+}
+
+// unsubscribe removes and closes the subscriber's channel. It holds the same
+// lock as Publish, so it never closes a channel mid-send.
+func (h *Hub) unsubscribe(id int) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	if ch, ok := h.subs[id]; ok {
+		delete(h.subs, id)
+		close(ch)
+	}
+}