diff --git a/.superpowers/sdd/carousel-wrap-report.md b/.superpowers/sdd/carousel-wrap-report.md
new file mode 100644
index 0000000..f832145
--- /dev/null
+++ b/.superpowers/sdd/carousel-wrap-report.md
@@ -0,0 +1,106 @@
+# Carousel wrap-around — implementation report
+
+## What changed
+
+### `main/modes/home.c`
+
+**`home_arrow_cb`** — replaced clamp logic with modular arithmetic:
+
+```c
+// Before (clamped):
+int idx = s_acur + dir;
+if (idx < 0) idx = 0;
+if (idx >= s_n_acts) idx = s_n_acts - 1;
+
+// After (wraps):
+int idx = (s_acur + dir + s_n_acts) % s_n_acts;
+```
+
+Adding `s_n_acts` before the modulo makes the `-1` direction safe (no
+negative modulo UB in C).
+
+**`home_scroll_cb`** — minor cleanup: `lv_obj_get_scroll_x` return type
+stored as `int32_t` before the division cast, then cast to `int` for the
+index. Clamp behaviour kept identical; swipe still snaps at edges.
+
+### `main/modes/games.c`
+
+Exactly parallel changes to `games_arrow_cb` and `games_scroll_cb` using
+`GAMES_N_TILES` instead of `s_n_acts`:
+
+```c
+// Before (clamped):
+int idx = s_gcur + dir;
+if (idx < 0) idx = 0;
+if (idx >= GAMES_N_TILES) idx = GAMES_N_TILES - 1;
+
+// After (wraps):
+int idx = (s_gcur + dir + GAMES_N_TILES) % GAMES_N_TILES;
+```
+
+## Arrow-wrap status
+
+**Fully working.** Both carousels wrap correctly on left/right arrow press:
+- pressing left on slide 0 jumps to slide N-1 (last)
+- pressing right on slide N-1 jumps to slide 0 (first)
+- `lv_obj_scroll_to_x` with `LV_ANIM_ON` is called with the correct target
+  x offset, and dots are updated immediately
+
+## Swipe-wrap approach
+
+**Clamped swipe (unchanged), arrow-only wrap implemented.**
+
+Infinite swipe wrap was investigated. The `LV_EVENT_SCROLL_END` callback
+fires after LVGL's snap animation has already settled, which means there
+is no clean hook point to silently teleport the scroll position. Calling
+`lv_obj_scroll_to_x(track, other_end, LV_ANIM_OFF)` from within
+`SCROLL_END` would work in principle but causes a visible jump (the track
+flashes to the other end) which is disorienting and fragile on slow
+hardware. The alternative approach of inserting sentinel/clone slides
+(infinite-scroll pattern) would require allocating two extra slides with
+duplicate LVGL objects, which costs PSRAM on an already tight ESP32-S3
+build.
+
+Decision: ship arrow-wrap only. For a children's device, the arrow buttons
+are the primary navigation affordance; swiping past the last slide is
+simply not possible (the content does not scroll further thanks to LVGL's
+natural scroll boundary).
+
+## Build result
+
+Full `idf.py build` could not be executed in the agent worktree because:
+1. The worktree has no `managed_components/` or pre-configured `build/`
+   tree (these live in the main checkout only).
+2. The main volume was 99% full (194 MB free), insufficient for a full
+   ESP-IDF build (typically >200 MB of incremental output).
+
+**Verification performed:** both files were compiled with `-fsyntax-only`
+using the exact flags extracted from the project's
+`build/compile_commands.json` (xtensa-esp32s3-elf-gcc, all ESP-IDF and
+managed-component include paths, LVGL defines). Both files compiled
+**without errors or warnings**.
+
+## Self-review
+
+- No `bsp_display_lock` added or removed — constraint respected.
+- `s_adots[8]` / `s_gdots[GAMES_N_TILES]` array bounds not touched.
+- No new allocations, no new LVGL objects.
+- Code style matches surrounding code (same spacing, same cast idioms).
+- `(cur + dir + n) % n` is the canonical safe modulo for both directions
+  in C (no negative operand entering the modulo).
+
+## Concerns
+
+1. **Swipe still clamps.** A user who swipes aggressively past slide 0 will
+   feel a hard wall. Acceptable for the target age group and consistent
+   with most children's apps.
+2. **`s_acur` / `s_gcur` state.** These static globals are initialised to 0
+   on first screen build and never reset. If the screens are rebuilt (they
+   are currently cached with `if (s_root) return s_root`), the saved
+   position is preserved correctly. If caching is ever removed, the globals
+   would need to be reset alongside the screen pointers.
+3. **Worktree vs main-tree build.** The commit is on
+   `worktree-agent-a66aef6ba538a372e` (which was reset to the tip of
+   `feat/routines`). The caller should cherry-pick or fast-forward into
+   `feat/routines`. A full `idf.py build` on the main checkout (which has
+   `managed_components/` and sufficient disk) will be needed before flashing.