diff --git a/.claude/worktrees/agent-a994a2ac1eff4f7dc b/.claude/worktrees/agent-a994a2ac1eff4f7dc
new file mode 160000
index 0000000..d2ab84d
--- /dev/null
+++ b/.claude/worktrees/agent-a994a2ac1eff4f7dc
@@ -0,0 +1 @@
+Subproject commit d2ab84d549fca95896a871b684b4b2c8cecbfb95
diff --git a/OPUS-BUILD-BRIEF-2.md b/OPUS-BUILD-BRIEF-2.md
index 2fd5982..955e087 100644
--- a/OPUS-BUILD-BRIEF-2.md
+++ b/OPUS-BUILD-BRIEF-2.md
@@ -64,26 +64,29 @@ Measured on the bench (repeat these before changing tuning):
 - Same press with The Best Thing → bites in 7 frames, 0 slips (sharp edges sink on downforce).
 - Wipe two passes → mass 0.51 → 0.09.
 
-**M11 is NOT finished — remaining before it can be called done:**
-1. No order actually routes through the bench yet. `Game.startPrep()` exists and the
-   harness drives it, but `Order.prep?: PrepStep[]` is declared and unused — no day
-   sends you there. Wire a day-10 prep order and play it end to end.
-2. `dice` two-phase and `wedges` are implemented in the sim but never driven —
-   no rotate-the-board input, no verification. M14 needs the dice; do it there or here.
-3. The pieces fan-out is placeholder-grade (cylinders). Pieces aren't physics bodies yet.
-4. The wipe is the same drag as cutting, disambiguated only by distance from the
-   ingredient — it needs a real cloth tool/mode before it reads to a stranger.
-5. No screenshots taken of the bench yet. Do that when the visuals are past placeholder.
+**UPDATE (2026-07-17, later): M11 is DONE** (`d7ce7b7`) — days 10/11 route through the
+bench (dice onion / wedge tomato), dice & wedges driven and measured (even: CV≈0 both;
+bad aim: CV 0.74; sloppy saw: mass 31 "a crime scene" — aim and tidiness are separate
+judged skills). Scorecard shows THE PREP → THE TOAST → THE SPREAD end to end.
+**M12 core is BUILT** (`sim/juicing.ts`, `scenes/juice.ts`, `dev-juice.ts` via `__tj`):
+verified — Pyramid 3revs@0.7 = 72.8% + 4 escaped pips, Juicernaut 100% + all seeds
+caught, butchered halve caps at its 0.6 theoretical (no double punishment), mash = 8.5%
++ pulp + forced seeds, reversals pay (77.1% vs 72.8% equal-sweep). NOT YET ROUTED:
+no order sends you to the juice station — integration point is `juiceResult()` /
+`juiceCriteria()` / `JUICE_LINES` in `sim/juicing.ts` (documented in-file). Remaining
+polish carried forward: pieces fan-out still cylinders (wrong for dice, non-blocking);
+wipe still distance-disambiguated, needs a cloth tool; MASTERPLAN.md now sits above
+this brief — read it first. Also in the cast since M11: mushrooms ×3 (slip showcase,
+`'mushroom'` silhouette), bell pepper (seed-scatter emergent win). See §7.5.
 
-**BLOCKER — the asset farm is down for 3D.** Every `bg_remove_local` job fails:
-m4pro's disk is 100% full (118MB free; ~3.4TB is `~/Music` — hands off, that's John's).
-Flux 2D still works — `bruschetta_hero.png` and `wonderslice_bag.png` generated fine and
-are committed. The whole 3D cast (bakery loaves, cutting board, prep-bench props) is
-stuck at the cutout stage. `mb settings` is owner-only (403), so re-routing the
-bg_remove lane needs John. Once there's space: `bash scripts/gen-assets.sh bakery`
-then `bash scripts/gen-assets.sh prep` — retry-once is now built into `gen_3d`, seeds
-are bumped, and the prep cast + 2D are already queued in the script. Everything
-degrades gracefully: `loadProp` catches, so the procedural stand-ins just stay.
+**Asset farm status (was BLOCKER, now half-fixed):** John freed 50GB on m4pro, but the
+cleanup took m4pro's venvs with it (`venvs/mflux`, `vendor/hunyuan3d-mlx/.venv`) — jobs
+routed to node m4 die in ~1s. `gen-assets.sh` now finds the client at `~/MODELBEAST`,
+reads `MB_TOKEN` from `~/Documents/backnforth/.env`, targets the m3ultra queue, and
+retries every stage 10× with 60s pauses so healthy nodes pick jobs up (ponytail-marked;
+strip once m4pro is rebuilt or pulled from the pool — owner-only, needs John). All 7
+bakery GLBs + the prep cast are landing; add mushroom/pepper/P2 prompts once the
+running batch finishes. Everything still degrades gracefully via `loadProp` catch.
 
 ---
 
diff --git a/public/assets/models/orange_whole.glb b/public/assets/models/orange_whole.glb
new file mode 100644
index 0000000..24c2929
Binary files /dev/null and b/public/assets/models/orange_whole.glb differ