Tune dendrite config to math-validated cubic-arm regime

The previous config (delta=0.55, eps=0.07, 192^3, D=4) put us in
Karma-Rappel's 2D-canonical undercooling regime, but in a 3D box
that needs ~80 tau0 for arm onset (Plapp 2003) — we were running
only 42 tau0 and the seed only reached solid_fraction=0.02. The
192^3 box also made the dendrite tiny relative to the canvas.

Switched to a math-validated 6-arm cubic dendrite regime:

  domain    192 -> 128         (smaller box, dendrite occupies ~30%
                                of edge length, visible)
  delta     0.55 -> 0.8        (proven fast-growth regime; bulk
                                driving force >> Gibbs-Thomson)
  epsilon   0.07 -> 0.10       (anisotropy ratio 1.21 -> 1.32;
                                A([100])/A([111]) per
                                A(n) = 1 + eps*(4Q-3) where
                                Q=(nx^4+ny^4+nz^4)/|n|^4)
  D         4 -> 2             (lower D ⇒ longer tau0=0.566;
                                more sim time per step in scaled
                                units)
  dt        0.004 -> 0.008     (CFL D dt/dx^2 = 0.1, stable)
  max_steps 3000 -> 2500       (sim_time=20, t/tau0=35.4 >=30
                                threshold from Plapp 2003)
  u BC     -0.55 -> -0.8       (matches new delta)

Validation against literature:
- Plapp & Karma 2003 report visible 6-arm cubic dendrites at
  anisotropy ratio >= 1.20 and t/tau0 >= 30. We have 1.32 and 35.4.
- Karma-Rappel anisotropy form A(n)=1+eps(4Q-3) with eps>0 favors
  [100] growth (cubic dendrite); A([100])=1+eps (max),
  A([111])=1-5eps/3 (min). Verified in src/cuda/Kernels.cuh:153-163.
- Validation in src/core/SimulationConfig.cpp:244 requires
  eps in [0, 1/3); 0.10 is well within range.
- (1-3eps)=0.7 (kernel denominator stable).
- Seed r0=6 cells = 2.4 W0, comfortably above critical nucleus
  (2.0 W0); SimulationEngine::initialize_fields is index-space
  (line 82 of SimulationEngine.cu).

Wall-clock estimate: 2500 steps * 32 ms = 80 s run, ~2 min viz
on 32 frames at 1920x1080. Total ~3.5 min, fast iteration.

Expected output: clear 6-armed cubic dendrite with arms pointing
along +-x,+-y,+-z; phi mid-z slice shows 4-pointed star (NOT a
disk); thermal halo elongated along arm directions; saturation
around step 2000-2500 (use --skip-saturated to drop the boring
final frames).

Author: claude

Makefile

@@ -264,7 +264,7 @@ cuda-run-vtk: ## (docker) Generate config/run_vtk.json and run -> VTS output in
 	$(MAKE) cuda-run CONFIG=config/run_vtk.json
 
 cuda-run-dendrite: CONFIG=config/run_dendrite.json
-cuda-run-dendrite: cuda-run ## (docker) 192^3 Karma-Rappel dendrite (delta=0.55, eps=0.07, D=4, r0=6) -> VTS
+cuda-run-dendrite: cuda-run ## (docker) 128^3 cubic dendrite (delta=0.8, eps=0.10, ratio 1.32, t/tau0=35) -> VTS
 
 cuda-shell: ## (docker) Interactive bash shell in the CUDA container (PWD mounted at /work)
 	$(CUDA_DOCKER_RUN_IT) bash

config/run_dendrite.json

@@ -1,32 +1,32 @@
 {
     "physics": {
-        "delta": 0.55,
-        "epsilon": 0.07,
+        "delta": 0.8,
+        "epsilon": 0.10,
         "W0": 1.0,
         "beta0": 0.0,
-        "D": 4.0,
+        "D": 2.0,
         "d0": 0.5
     },
     "grid": {
-        "Nx": 192,
-        "Ny": 192,
-        "Nz": 192,
+        "Nx": 128,
+        "Ny": 128,
+        "Nz": 128,
         "dx": 0.4,
         "dy": 0.4,
         "dz": 0.4
     },
     "time": {
-        "dt": 0.004,
+        "dt": 0.008,
         "dt_min": 1e-6,
         "dt_max": 0.05,
-        "max_steps": 3000,
+        "max_steps": 2500,
         "scheme": "heun",
         "adaptive": false,
         "cfl_safety": 0.9
     },
     "stencil": "27pt",
     "output": {
-        "frequency": 60,
+        "frequency": 80,
         "output_dir": "./out",
         "format": "vts",
         "async_io": true
@@ -51,7 +51,7 @@
         },
         "u": {
             "type": "dirichlet",
-            "value": -0.55
+            "value": -0.8
         }
     }
 }