diff --git a/.gitea/workflows/dev-deploy.yaml b/.gitea/workflows/dev-deploy.yaml
index 5df15df..2a2ef6b 100644
--- a/.gitea/workflows/dev-deploy.yaml
+++ b/.gitea/workflows/dev-deploy.yaml
@@ -148,14 +148,80 @@ jobs:
             -v "${{ gitea.workspace }}/pkg/geoip/test-data/test-data:/src:ro" \
             alpine sh -c 'cp /src/GeoIP2-Country-Test.mmdb /dst/geoip.mmdb'
 
+      - name: Recycle engine containers on image drift
+        run: |
+          # Compare the freshly-built `galaxy-engine:dev` SHA against
+          # every running `galaxy-game-*` container. The backend
+          # reconciler adopts pre-existing labelled engine containers
+          # without checking image drift, so a running sandbox would
+          # otherwise keep serving the previous engine code until the
+          # container is recycled by hand. This step makes the recycle
+          # automatic but only when it is actually needed:
+          #
+          #   * BuildKit cache hit on the `Build galaxy-engine image`
+          #     step → `galaxy-engine:dev` keeps its previous SHA →
+          #     no drift → no-op (no engine source change to deploy).
+          #   * engine source change → fresh SHA → for each drifted
+          #     container we stop the backend, remove the container,
+          #     wipe its bind-mounted state directory (Engine.Init()
+          #     writes turn-0 over any pre-existing `turn-N` files —
+          #     silent state corruption otherwise), and cascade-delete
+          #     the lobby `games` row (the FKs in `00001_init.sql`
+          #     drop the matching `runtime_records`, `memberships`,
+          #     `player_mappings`, etc. in the same write). The
+          #     `dev-sandbox` bootstrap on the next backend boot finds
+          #     no live sandbox and provisions a fresh one on the new
+          #     engine image.
+          #
+          # Backend is stopped first to keep the reconciler from
+          # racing the recycle (mid-stream adoption / restart). The
+          # subsequent `Bring up the stack` step restarts it.
+          set -u
+          new_sha=$(docker image inspect galaxy-engine:dev --format '{{.Id}}')
+          echo "fresh galaxy-engine:dev = $new_sha"
+
+          drift=()
+          for c in $(docker ps --filter "name=galaxy-game-" --format '{{.Names}}'); do
+            cur=$(docker inspect "$c" --format '{{.Image}}')
+            if [ "$cur" != "$new_sha" ]; then
+              drift+=("${c#galaxy-game-}")
+              echo "  drift: $c was on $cur"
+            else
+              echo "  match: $c"
+            fi
+          done
+          if [ ${#drift[@]} -eq 0 ]; then
+            echo "no drift detected — recycle skipped"
+          else
+            docker stop -t 30 galaxy-dev-backend >/dev/null 2>&1 || true
+            state_root="$HOME/.galaxy-dev/game-state"
+            for gid in "${drift[@]}"; do
+              echo "recycling $gid"
+              docker rm -f "galaxy-game-$gid" >/dev/null 2>&1 || true
+              # Wipe the per-game state dir as root inside a throwaway
+              # container so we can remove files left behind by the
+              # engine container even when its uid differs from the
+              # runner's.
+              docker run --rm -v "$state_root:/state" alpine \
+                sh -c "rm -rf -- /state/$gid"
+            done
+            ids_csv=$(printf "'%s'," "${drift[@]}")
+            ids_csv=${ids_csv%,}
+            docker exec galaxy-dev-postgres psql -v ON_ERROR_STOP=1 \
+              -U galaxy -d galaxy_backend \
+              -c "DELETE FROM backend.games WHERE game_id IN (${ids_csv});"
+          fi
+
       - name: Reap stray dev-deploy containers
         run: |
           # Remove any non-running compose-managed containers from
           # earlier deploys before `compose up`. Filter by the stack
           # label so we never touch unrelated workloads on the same
-          # daemon. Running containers (incl. engine instances backend
-          # spawned itself with the same label) are left intact —
-          # those are reattached by the backend reconciler on boot.
+          # daemon. Running engine containers spawned by backend with
+          # the same label are left intact when their image SHA still
+          # matches the freshly-built `galaxy-engine:dev` (handled by
+          # the preceding `Recycle engine containers on image drift`
+          # step); the reconciler reattaches them on backend boot.
           ids=$(docker ps -aq \
             --filter "label=galaxy.stack=dev-deploy" \
             --filter "status=exited" \
diff --git a/tools/dev-deploy/README.md b/tools/dev-deploy/README.md
index 8b3fd13..2a7a9fe 100644
--- a/tools/dev-deploy/README.md
+++ b/tools/dev-deploy/README.md
@@ -235,6 +235,29 @@ The deploy is idempotent — when the PR later merges into
 healthcheck steps, overwriting whatever the manual dispatch left
 behind. There is no separate state to clean up between the two paths.
 
+### Engine image drift recycle
+
+`backend` spawns one engine container per game (the long-lived "Dev
+Sandbox" plus any user-created games) and the reconciler reattaches
+to whatever it finds with the `galaxy.stack=dev-deploy` label. That
+reattach does not check the running container's image SHA against the
+freshly-built `galaxy-engine:dev` tag, so an unchanged container would
+otherwise keep serving the previous engine code after a redeploy.
+
+The `dev-deploy.yaml` workflow handles this in the
+`Recycle engine containers on image drift` step. When `docker build`
+produces a new `galaxy-engine:dev` SHA, the step compares it against
+every running `galaxy-game-*` container and, for each drifted one,
+stops the backend, removes the container, wipes its bind-mounted
+state directory (Engine.Init() writes turn-0 over any pre-existing
+`turn-N` files), and cascade-deletes the lobby `games` row. The
+`dev-sandbox` bootstrap on the next backend boot finds no live
+sandbox and provisions a fresh one on the new engine image.
+
+When the engine sources are unchanged, the BuildKit cache hits and
+the SHA stays the same — the recycle step is a no-op and the running
+games keep their state across the deploy.
+
 ## Relationship to other infrastructure
 
 - `tools/local-dev/` — single-developer playground, host-port mapped,