local-dev: rebuild dead sandbox + harden lobby card UX

Three fixes around the dev sandbox end-to-end path. Each one was
flushed out by an actual login walkthrough after the previous
commit.

Backend bootstrap now treats `cancelled`, `finished`, and
`start_failed` as terminal: the per-boot find-or-create skips such
games and provisions a fresh one. Without this, a single bad
shutdown cascade leaves the developer staring at a dead lobby tile
forever (cancelled games don't transition back). Covered by
TestTerminalSandboxStatus.

Tools/local-dev: stop killing engine containers in `make down`. The
runtime treats the disappearance of an engine as a real failure
(cascading the lobby game to `cancelled`); leaving the container
running across `down/up` lets the runtime reconciler re-attach on
the next boot. The teardown happens only in `make clean`, where the
DB is wiped anyway. Compose now also exposes :9090 (authenticated
EdgeGateway listener) on the host so the Vite dev proxy can reach
the Connect-Web surface, and bumps the gateway anti-abuse limits
for `public_misc` so the same surface is not blanket-rejected with
413.

Ui/frontend: the lobby's `My Games` cards are now clickable only
for the playable statuses (`running`, `paused`, `finished`). All
other statuses render as disabled buttons so a click on a draft or
cancelled game no longer drops the user on a 404 — the in-game
view at /games/:id/* doesn't exist before Phase 10 and never makes
sense for a cancelled game. Vite proxy splits the dev targets so
`/api/*` continues to talk to the REST listener and
`/galaxy.gateway.v1.EdgeGateway/*` is routed to the Connect-Web
listener via VITE_DEV_GRPC_PROXY_TARGET (defaults to :9090).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

tools/local-dev/Makefile

@@ -13,11 +13,11 @@ ENGINE_LABEL := org.opencontainers.image.title=galaxy-game-engine
 help:
 	@echo "Local development stack for the Galaxy UI:"
 	@echo "  make up             Build (if needed) and bring up the stack, wait until healthy"
-	@echo "  make down           Stop containers (incl. spawned engines), keep volumes"
+	@echo "  make down           Stop compose containers, leave engines + volumes intact"
 	@echo "  make rebuild        Force rebuild of backend / gateway images and bring up"
 	@echo "  make build-engine   Build the engine image $(ENGINE_IMAGE) used by the dev sandbox"
 	@echo "  make stop-engines   Stop and remove only the per-game engine containers"
-	@echo "  make clean          Stop and wipe volumes (postgres data, engines, game state)"
+	@echo "  make clean          Stop everything (incl. engines) and wipe volumes + game state"
 	@echo "  make logs           Tail all logs"
 	@echo "  make logs-backend   Tail only the backend logs"
 	@echo "  make logs-gateway   Tail only the gateway logs"
@@ -47,7 +47,7 @@ build-engine:
 		docker build -t $(ENGINE_IMAGE) -f $(REPO_ROOT)/game/Dockerfile $(REPO_ROOT); \
 	fi
 
-down: stop-engines
+down:
 	$(COMPOSE) down
 
 clean: stop-engines
@@ -58,9 +58,11 @@ clean: stop-engines
 	fi
 
 # Spawned engine containers run outside the compose project (the
-# backend's runtime creates them on demand), so `compose down` does
-# not see them. We discover them by the engine image's
-# OCI title label, set by game/Dockerfile.
+# backend's runtime creates them on demand). They intentionally
+# survive `make down` so the runtime reconciler can reattach on the
+# next `make up` — killing them out of band makes the runtime
+# cascade the game to `cancelled`. We only remove them as part of
+# `clean`, where the whole DB is wiped anyway.
 stop-engines:
 	@ids=$$(docker ps -aq --filter label=$(ENGINE_LABEL)); \
 	if [ -n "$$ids" ]; then \