Fable (integrator): round-3 update — publish authorized, hero props go local

John authorized origin push (done) + depot publish (he runs it; GOD3_PW stays out of
agent hands). Hero props switch from the gated $2.70 fal.ai batch to the free
MODELBEAST path: sf3d + TRELLIS.2 weights verified already cached on-device (the
round-2 'HF-gated, setup needed' note was stale) — flux_local concepts -> trellis_mac
GLBs -> normalize.py, fal.ai only as per-reject fallback.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/LANES/ROUND3_INSTRUCTIONS.md

@@ -19,8 +19,13 @@ validation, and the last two integration steps** (LANE_F_NOTES §3.4/§3.5).
 2. **True closed residential ring → v2** (stays in V2_IDEAS).
 3. **F finding #4 (doubled possessives) is CLOSED** — verified 0 across 2,418 shop names /
    5 seeds on current main. It predated Lane A's round-1 fix.
-4. **Depot publish and the ~$2.70 MeshGod hero batch remain gated on John** — do not publish,
-   do not spend. Local validation paths below unblock everything else.
+4. **John has authorized the depot publish** (2026-07-14). It needs `GOD3_PW` in the env, which
+   agents don't have and must not hunt for — John runs `pipeline/publish.py` himself (dry-run
+   verified: 16 GLBs + thumbs ready). After it lands, Lane F adds `--depot` to the QA gate.
+5. **Hero props: free-first on MODELBEAST — the local path is READY, no setup needed.** The
+   round-2 "HF-gated, setup needed" note is stale: `sf3d` (3.7 GB) and `TRELLIS.2-4B` (14 GB)
+   weights are already cached on this box with working venvs. Do not spend the ~$2.70 fal.ai
+   batch up front; it's the fallback for local rejects only.
 
 ## Process rules (unchanged, they matter)
 
@@ -84,18 +89,27 @@ in `pipeline/_normalized/` — you don't need the depot publish.
    chunk build/dispose) — design only, no implementation; it's the v1.5 fix for big-town
    street emptiness.
 
-## Lane E (Assets) — local depot mode + near-misses
+## Lane E (Assets) — hero props on MODELBEAST + local depot mode
 
-Publish stays gated (needs `GOD3_PW` + provenance recording — John's call).
-1. **Unblock local validation**: make `core/loaders.js` (or a manifest/URL flag, e.g.
+Publish is **authorized** but John runs it himself (`GOD3_PW`): record the per-asset
+provenance list in AUDIT.md first, then hand him the two-liner. Your new headline task:
+1. **Hero props, free, on-device.** The `meshgod_batch.json` gap list (glass case, fridge,
+   till, listening booth, …) now goes through MODELBEAST instead of fal.ai:
+   `gen_skins.py`-style concept images via `flux_local` (proven on the facades) →
+   `trellis_mac` image→GLB (`1024_cascade`, ~3–5 min/gen; use `sf3d` for seconds-fast drafts
+   to iterate the concept image before committing to a TRELLIS run) → `normalize.py` house
+   law → thumbnails → manifest. Weights are already cached (verified 2026-07-14); both
+   operator venvs exist under `~/Documents/MODELBEAST/server/operators/`. **Eyeball every
+   thumbnail**; anything unusable after 2 local attempts goes on the fal.ai fallback list
+   (still gated, ask John — budget ≤$2.70 total).
+   First check the DealGod cache / depot for reuse (cash-register + glass-case may exist).
+2. **Unblock local validation**: make `core/loaders.js` (or a manifest/URL flag, e.g.
    `?localdepot=1`) resolve `depot:<file>` to the locally staged GLBs so Lanes B/C/F can
-   validate before publish. Keep the placeholder-on-miss behaviour.
-2. Support Lane C's visual validation — fix any origin/footprint mismatch they report
+   validate before/without the depot. Keep the placeholder-on-miss behaviour.
+3. Support Lane C's visual validation — fix any origin/footprint mismatch they report
    (your normalize.py now handles hierarchies; re-run per asset).
-3. Optional: manual Blender re-origin pass on the 3 near-misses (garage_shelves, radiogram,
+4. Optional: manual Blender re-origin pass on the 3 near-misses (garage_shelves, radiogram,
    market_stall) if C wants more variety; otherwise leave to primitives.
-4. Have the publish command + per-asset provenance list **staged and documented** so the
-   moment John says go it's a 5-minute job.
 
 ## Lane F (Integration) — lead lane this round
 
@@ -116,8 +130,13 @@ Priority order, per your own §3 plan:
 
 ## Needs a human (John)
 
-- **Push to origin** — main is now 10 commits ahead of origin/main (only the scaffold is pushed).
-- **Depot publish go/no-go** (needs `GOD3_PW`; E will have it staged).
-- **~$2.70 MeshGod hero-prop batch** go/no-go (pure polish; primitives cover everything).
-- **A real pointer-lock playtest** — automation can't engage pointer lock;
+- ~~Push to origin~~ — **done** (2026-07-14, Fable).
+- **Run the depot publish** (authorized; agents don't hold the cred):
+  ```bash
+  cd /Users/m3ultra/Documents/procity
+  export GOD3_PW=<the depot password>
+  python3 pipeline/publish.py   # idempotent; dry-run already verified 16 GLBs + thumbs
+  ```
+- **fal.ai spend**: only if MODELBEAST hero-prop rejects need it (≤$2.70, per-prop go).
+- **A real pointer-lock playtest** (after the current Opus round, per John) —
   `cd web && python3 -m http.server 8130` → http://localhost:8130?seed=20261990.