Native magma --rl-bin transfer eval (closed and replay)

AGENTS.md

@@ -1,7 +1,7 @@
 # netherite (AGENTS.md)
 
-Home: **Anvil-primary** - canonical at `anvil:~/dev/netherite`. Build and run
-here. MacBook is control plane / Moonlight / image viewing only.
+Home: **Anvil-primary** - canonical at `anvil:~/dev/netherite`. Build, run,
+and validate here. MacBook is control plane / Moonlight / image viewing only.
 
 Git remote: **https://github.com/Infatoshi/netherite**. That is the only
 GitHub repo. `origin` on both machines points there.
@@ -14,10 +14,20 @@ next to the code they govern.
 
 | OS | Role |
 |----|------|
-| **Linux x86_64** | Full stack. Build CPU/CUDA, run Java oracle, train blaze, sweep. Canonical host: anvil (Ubuntu). Needs JDK 8 + NVIDIA CUDA for GPU paths. |
+| **Linux x86_64** | Full stack. Build CPU/CUDA, run Java oracle, train blaze, sweep. Canonical host: anvil (Ubuntu, 9950X3D). Needs JDK 8 + NVIDIA CUDA for GPU paths. All Linux validation runs here. |
 | **macOS** | Build and verify the CPU and Metal game backends. Also serves as the control plane and image/video review host. Do not expect native `runClient` or CUDA here. |
 | **Windows** | Not a supported build/run host for this monorepo. Use WSL2 Linux if you must, or a remote Linux box. |
 
+### Hosts (where work runs)
+
+| host | Role |
+|------|------|
+| **anvil** | Canonical clone. Linux validation: `make test`, magma CPU, tape replay, blaze M1/M2, `eval` (cpu/cuda/magma), oracle capture. CUDA `sm_120`. Shared GPUs: `nvidia-smi` first; long GPU jobs through `overnight-compute`. |
+| **macOS** | Control plane. Metal build, `cpu==metal` kernel half, image/video review. Not the Linux validation host. |
+| **gamer** | Not the validation host. Use only when a job needs the 3090 / `sm_86`. |
+
+Do not stage magma CPU gates, 13-seed eval, tape replay, or `make test` on gamer. Anvil is the faster CPU.
+
 Prism / MultiMC / official launcher: optional jar source for assets. Fresh
 boxes do **not** need Prism credentials; `make -C java bootstrap-oracle` pulls
 MC 1.11.2 via ForgeGradle (you must own the game).
@@ -65,9 +75,10 @@ sub-agents pointed at them without waiting for a human to pick items:
 1. Pick from `magma/OPEN_DIVERGENCES.md` class A (top down) and
    `blaze/OPEN_DIVERGENCES.md` unported rows (dependency order). One lane
    per item. Skip items whose evidence (tape, golden) is not on this Mac.
-2. Stage: `bash scripts/lane_stage.sh <lane> <host> [--tape NAME]...`
-   (gamer for magma CPU work; anvil for oracle captures, CUDA, blaze M2;
-   Metal half of a twins lane builds on the Mac).
+2. Stage: `bash scripts/lane_stage.sh <lane> anvil [--tape NAME]...`
+   (anvil for magma CPU, oracle captures, CUDA, blaze M1/M2, eval.
+   Metal half of a twins lane builds on the Mac. Do not stage
+   validation on gamer.)
 3. Prompt = `docs/SUBAGENT.md` + a goal block: the item text, the gate
    command, the documented baseline numbers, the hard goal, and what is
    forbidden. Launch from the lane worktree, in the background.

blaze/OPEN_DIVERGENCES.md

@@ -16,6 +16,7 @@ Definitions, so the list stays honest:
 - Gate runner: `blaze/env/port_matrix.py` over `blaze/env/port_matrix.yaml`
   (fail-closed; VERIFIED / BLOCKED / FAILED per row and tier).
 
+Last documented: 2026-08-26 magma 13-seed transfer (`retrain_0821_best.bin` closed == cpu == cuda, 0/13 torches; replay MATCH 56/65. `ppo_ckpt_best.bin` closed t0:13; replay MATCH 45/65. Cam-px DIVERGE not blessed). No new M1/M2.
 Last verified: lane/blocklight 2026-08-25 (Sweep row 3: World.checkLightFor BLOCK flood. test_blocklight matches magma Manhattan 0..14. Fluids M2 raw/warp/scalar VERIFIED after moving CU_LIGHT_Q off the CUDA thread stack. listed --no-deps M1 27 VERIFIED; M2 27 VERIFIED; mining_slice M1+M2 BLOCKED rc=3. Root make test PASS. SNAP v11 unchanged. No BP_ light digest.)
 Last verified: lane/potions 2026-08-24 (Sweep row 13: SNAP v11 potion trailer after resume v10. Magma skip double fire decrement on shared hostile path. potions M1+M2 VERIFIED 120 raw/warp/scalar + resume N=90 M=30 dump version=11. listed --no-deps M1 27 VERIFIED; M2 27 VERIFIED; mining_slice BLOCKED: s14_t0_r48_no_liquid.bsnp is a v1 bake, 12 *_d*.bsnp present. Root make test PASS.)
 Last verified: lane/resumegate 2026-08-23 (Sweep rows 4 and 8: snapshot v10 resume trailers + mob sidecars. continuous-vs-resume BP_ gate. listed --no-deps M1 VERIFIED; M2 VERIFIED raw/warp/scalar except mining_slice BLOCKED missing *_d*.bsnp).
@@ -122,15 +123,55 @@ Two consequences worth stating plainly:
   48-point path cap), IntHashMap aliasing reproduced, all 8 detmob tapes
   must be blaze-exact.
 
+## Spawn -> dragon (policy vs magma)
+
+A policy in blaze CUDA cannot complete magma `PRODUCT.md` steps 3-9
+(Nether portal, fortress/rods, eyes, stronghold, End, dragon, `won`)
+while matching magma. Gate 2 accepts spawn->torch. `docs/GATES.md`
+forbids dragon-fight RL. Isolated kernels (`nether_portal.h`,
+`end_portal.h`, `nether_full.h`, `chunk_provider_nether.h`,
+`ender_dragon.h`) are CPU==CUDA units; they are not `blaze_tick`.
+Java-vs-magma leftovers on that same route stay in
+`magma/OPEN_DIVERGENCES.md` (stronghold L, tick order M, detmob nether
+A*, class A pixels). Do not file missing Nether/End in this env as a
+magma bug.
+
+| PRODUCT step | Blaze env | Where |
+|---|---|---|
+| Overworld spawn, dig, craft, furnace, chests, fluids, TNT, overworld hostiles | yes, M1+M2 inside a `.bsnp` | Verified rows |
+| Food / bed / buckets / flint / portal ignite | buckets tick but are not in `rl_crafts`; fire->portal follower not ported; `dimension` is always 0; id 51 edits unreachable | `blaze_core.h` (`cu_world_set_state` follower); sweep 2 |
+| Linked Nether / End | no | unported `portals_dimensions`, `nether_route` |
+| Fortress, blaze, ghast, pigman, magma cube, wither skeleton, silverfish | type tags exist (`entity_hostile_spine.h`); spawn table and live AI in the env are overworld only | `hostile_spawn.h` `hs_to_ew`; `blaze_core.h` header |
+| Throw eyes of ender / fireballs | magma-only | projectiles close note |
+| Stronghold hunt, end-portal frames, enter End | no dimension in the env. Magma vs Java stronghold is magma OPEN sweep 1 | magma `OPEN_DIVERGENCES.md` |
+| Dragon, crystals, bed explode in DIM -1/1, exit portal, `won` | `ender_dragon.h` is a 200-tick fixed arena (no block destruction, no contact damage), not wired into `blaze_tick`. Bed explode in Nether/End is magma-only | unported `dragon_victory`; `player_bed.h` |
+
+Even on overworld snapshots, a trained policy does not see or act as
+magma `player_ctl`:
+
+- Sweep 1: episode region is the snapshot (`cu_region_idx`,
+  `cu_recenter`); magma streams chunks. L.
+- Sweep 2: no strafe, sneak, sprint, or slot-click. Camera planes are
+  log / leaves / coal / stone / dirt / table / occupancy / depth /
+  edge (`obs_pack.h` `pack_frame`). No mob, light, or health plane.
+- Sweep 6: death is terminal.
+- Detmob A* is magma-CPU (above).
+- M1/M2 are BP_ digests inside that region, not a streamed survival
+  world.
+
+Closing the port-matrix DAG does not by itself make a dragon policy:
+sweep 1 and 2 still block long-horizon transfer. An isolated kernel
+PASS is not env support.
+
 ## Sweep 2026-08-23 (codex full read, magma -> blaze; unverified by gate unless noted)
 
 Ranked by RL fidelity. Each row: magma site; blaze site or absent; what
 closing needs. Spot-checked by hand on 2026-08-23: rows 7 and 10 confirmed.
 
-1. Episode region is fixed; no chunk streaming (`blaze/env/blaze_core.h:
-   456-516, 4652-4659`). Magma streams chunks around the player. Every
-   row is VERIFIED only inside the snapshot region. L; blocks long-horizon
-   transfer.
+1. Episode region is fixed; no chunk streaming (`cu_region_idx`,
+   `cu_recenter` in `blaze/env/blaze_core.h`). Magma streams chunks around
+   the player. Every row is VERIFIED only inside the snapshot region. L;
+   blocks long-horizon transfer. See "Spawn -> dragon" above.
 2. Policy actions are privileged helpers (craft/interact/smelt in
    `blaze/rl/obs_pack.h:20-51`); no strafe, sneak, sprint, or slot-click
    as magma `player_ctl` sees them. `do_place` is reachable: RL `a[8]` is
@@ -228,8 +269,14 @@ Observation exposes no mob, light, or health planes.
 
 - Native ppo spawn->torch t0 is 0.215 (chain4 curriculum, 510M ticks) vs the
   0.4 target; stage4->torch is 8/8 seeds. GATES row 1.
-- No native transfer/eval of `ppo_ckpt.bin` into magma (torch eval scripts
-  removed). GATES row 2.
+- Native magma eval is wired (`eval --backend magma --transfer closed|replay`)
+  and the 32-tick BOLR harness is `make -C blaze/rl test-eval-magma`.
+  13-seed tries=5 n=65 (gamer 2026-08-26): `retrain_0821_best.bin` magma
+  closed == cpu == cuda (torches 0/13, t0:6 logs3:7). Replay MATCH 56/65;
+  9 DIVERGE all cam 1-5 px, not blessed. `ppo_ckpt_best.bin` magma
+  closed == cpu == cuda t0:13; replay MATCH 45/65. Camera stays compile-time 64x36
+  (`oc_pixel`); `--rl-bin` is not a resolution flag. GATES row 2. Gate 2
+  accept stays open (0 torches).
 - Python still owns replay/pixels/M2 verify; no binary tape; no root
   `make verify`. GATES row 13.
 - Blaze native Metal tick (M3) is sequenced after CUDA survival rows pass

blaze/rl/Makefile

@@ -110,6 +110,9 @@ TEST_SKY_BIN := $(OUT)/test_skylight_water
 TEST_BLK_BIN := $(OUT)/test_blocklight
 SMOKE_BIN := $(OUT)/ppo
 EVAL_BIN := $(OUT)/eval
+EVAL_MAGMA_OBJ := $(OUT)/eval_magma.o
+TEST_EVAL_MAGMA_BIN := $(OUT)/test_eval_magma
+CKPT ?= out/blaze/rl/ppo_ckpt_best.bin
 SAN_TEST_BIN := $(SAN_OUT)/test_train_config
 SAN_CHAIN_BIN := $(SAN_OUT)/test_chain_reward
 SAN_SMOKE_BIN := $(SAN_OUT)/ppo
@@ -122,7 +125,7 @@ CHAIN_OBJS := $(OUT)/chain_reward.o $(OUT)/chain_curr.o
 .PHONY: all test-config test-chain test-capture test-entity-spine \
 	test-falling test-weather test-chests test-furnaces test-projectiles test-explosions test-randtick test-biome-plane test-mobs test-mobs-ss test-mobs-end test-mobs-witch test-passives test-hazards test-ground-items test-potions test-placement test-beds test-skylight-water test-blocklight smoke smoke-cpu smoke-cuda \
 	smoke-metal smoke-reset test sanitize clean env-cpu env-cuda \
-	env-metal eval
+	env-metal eval test-eval-magma eval-magma
 
 all: test-config smoke
 
@@ -147,6 +150,9 @@ $(OUT)/chain_reward.o: chain_reward.c chain_reward.h | $(OUT)
 $(OUT)/chain_curr.o: chain_curr.c chain_curr.h chain_reward.h | $(OUT)
 	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -c -o $@ chain_curr.c
 
+$(EVAL_MAGMA_OBJ): eval_magma.c eval_magma.h | $(OUT)
+	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -c -o $@ eval_magma.c
+
 # ---- Linux CUDA staging ----
 ifneq ($(UNAME_S),Darwin)
 $(OUT)/env_cuda_stage.o: env_cuda_stage.c env_cuda_stage.h | $(OUT)
@@ -400,21 +406,22 @@ $(SMOKE_BIN): ppo.c obs_pack.h rl_ckpt.h $(TRAIN_CFG_OBJ) $(CHAIN_OBJS) \
 		$(CHAIN_OBJS) $(METAL_OBS_OBJ) $(METAL_HOST_OBJ) $(NN_OBJS) \
 		$(LINK_LIBS) $(LDLIBS)
 
-$(EVAL_BIN): eval.c obs_pack.h rl_ckpt.h $(CHAIN_OBJS) $(METAL_OBS_OBJ) \
-		$(METAL_HOST_OBJ) $(NN_OBJS) $(ENV_CPU_SO) $(METAL_LIB) | $(OUT)
-	$(METAL_CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ eval.c $(CHAIN_OBJS) \
-		$(METAL_OBS_OBJ) $(METAL_HOST_OBJ) $(NN_OBJS) \
+$(EVAL_BIN): eval.c obs_pack.h rl_ckpt.h $(CHAIN_OBJS) $(EVAL_MAGMA_OBJ) \
+		$(METAL_OBS_OBJ) $(METAL_HOST_OBJ) $(NN_OBJS) $(ENV_CPU_SO) \
+		$(METAL_LIB) | $(OUT)
+	$(METAL_CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ eval.c $(EVAL_MAGMA_OBJ) \
+		$(CHAIN_OBJS) $(METAL_OBS_OBJ) $(METAL_HOST_OBJ) $(NN_OBJS) \
 		$(LINK_LIBS) $(LDLIBS)
 else
 $(SMOKE_BIN): ppo.c obs_pack.h rl_ckpt.h $(TRAIN_CFG_OBJ) $(CHAIN_OBJS) \
 		$(STAGE_OBJ) $(NN_OBJS) $(ENV_CPU_SO) $(ENV_CUDA_SO) | $(OUT)
 	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ ppo.c $(TRAIN_CFG_OBJ) $(CHAIN_OBJS) \
 		$(STAGE_OBJ) $(NN_OBJS) $(LINK_LIBS)
 
-$(EVAL_BIN): eval.c obs_pack.h rl_ckpt.h $(CHAIN_OBJS) $(STAGE_OBJ) $(NN_OBJS) \
-		$(ENV_CPU_SO) $(ENV_CUDA_SO) | $(OUT)
-	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ eval.c $(CHAIN_OBJS) $(STAGE_OBJ) \
-		$(NN_OBJS) $(LINK_LIBS)
+$(EVAL_BIN): eval.c obs_pack.h rl_ckpt.h $(CHAIN_OBJS) $(EVAL_MAGMA_OBJ) \
+		$(STAGE_OBJ) $(NN_OBJS) $(ENV_CPU_SO) $(ENV_CUDA_SO) | $(OUT)
+	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ eval.c $(EVAL_MAGMA_OBJ) \
+		$(CHAIN_OBJS) $(STAGE_OBJ) $(NN_OBJS) $(LINK_LIBS)
 endif
 
 test-config: $(TEST_BIN)
@@ -556,6 +563,19 @@ smoke-reset: $(SMOKE_BIN)
 # and t0 snaps). CUDA/Metal slot via --backend; default is CPU.
 eval: $(EVAL_BIN)
 
+# 32-tick Magma --rl-bin vs blaze_cpu BOLR. Not on `test`. Needs magma_game.
+$(TEST_EVAL_MAGMA_BIN): test_eval_magma.c $(EVAL_MAGMA_OBJ) $(ENV_CPU_SO) | $(OUT)
+	$(CC) $(OWN_CFLAGS) $(CPPFLAGS) -o $@ test_eval_magma.c \
+		$(EVAL_MAGMA_OBJ) $(LDLIBS)
+
+test-eval-magma: $(TEST_EVAL_MAGMA_BIN) $(ENV_CPU_SO)
+	cd $(REPO) && ./out/blaze/rl/test_eval_magma
+
+# 13-seed magma closed-loop. Override CKPT=. Not on `test`.
+eval-magma: $(EVAL_BIN) $(ENV_CPU_SO)
+	cd $(REPO) && ./out/blaze/rl/eval --checkpoint $(CKPT) \
+		--backend magma --stage 0
+
 test: test-config test-chain test-capture smoke
 
 # Sanitizer build of owned sources; env via dlopen of the same .so files.