refactor: 将子模块转为普通目录，移除外部 git 依赖

- 移除 JEPA/lejepa-identifiability 子模块 gitlink
- 移除 research/multiply/MultiPLY 子模块 gitlink
- 删除 .gitmodules（不再有外部 URL 依赖）
- 两个目录内容作为普通文件纳入主仓库追踪
- 删除各自内部 .git 目录，消除嵌套 git 仓库

JEPA/lejepa-identifiability/experiments/configs/2d.yaml

@@ -0,0 +1,30 @@
+# 2D illustration + Gaussian half of the regularizer ablation
+# python run.py --config configs/2d.yaml --run spiral_lejepa --seed 1337
+
+experiment: 2d
+out: results/2d
+
+# Data
+N: 2
+source_dist: gaussian
+num_eval: 10000
+
+# Training (shared)
+steps: 20000
+lr: 3.0e-3
+batch_size: 256
+rho: 0.95
+log_every: 500
+
+# Per-run specs: 4 mixings x 2 objectives
+runs:
+  spiral_lejepa:   {mixing: spiral,   encoder: mlp,     hidden: 256, mode: lejepa, lamb: 1.0e-3}
+  spiral_whiten:   {mixing: spiral,   encoder: mlp,     hidden: 256, mode: whiten, lamb: 0.5}
+  banana_lejepa:   {mixing: banana,   encoder: mlp,     hidden: 256, mode: lejepa, lamb: 1.0e-3}
+  banana_whiten:   {mixing: banana,   encoder: mlp,     hidden: 256, mode: whiten, lamb: 0.5}
+  sinusoid_lejepa: {mixing: sinusoid, encoder: mlp,     hidden: 256, mode: lejepa, lamb: 1.0e-3}
+  sinusoid_whiten: {mixing: sinusoid, encoder: mlp,     hidden: 256, mode: whiten, lamb: 0.5}
+  nvp_lejepa:      {mixing: nvp,      encoder: matched, n_layers: 8, mode: lejepa, lamb: 1.0e-3}
+  nvp_whiten:      {mixing: nvp,      encoder: matched, n_layers: 8, mode: whiten, lamb: 0.5}
+
+seeds: [1337, 1338, 1339]

JEPA/lejepa-identifiability/experiments/configs/gennorm.yaml

@@ -0,0 +1,32 @@
+# Generalized normal sweep across mixings (main-text figure)
+# python run.py --config configs/gennorm.yaml --run spiral_lejepa --alpha 2.0 --seed 1337
+
+experiment: gennorm
+out: results/gennorm
+
+N: 2
+source_dist: gennorm   # alpha provided per-run via CLI
+num_eval: 10000
+
+steps: 20000
+lr: 3.0e-3
+batch_size: 256
+rho: 0.95
+log_every: 500
+
+runs:
+  spiral_lejepa:    {mixing: spiral,   encoder: mlp,     hidden: 256, mode: lejepa,  lamb: 1.0e-3}
+  spiral_whiten:    {mixing: spiral,   encoder: mlp,     hidden: 256, mode: whiten,  lamb: 0.5}
+  spiral_infonce:   {mixing: spiral,   encoder: mlp,     hidden: 256, mode: infonce, sigma: 1.0}
+  banana_lejepa:    {mixing: banana,   encoder: mlp,     hidden: 256, mode: lejepa,  lamb: 1.0e-3}
+  banana_whiten:    {mixing: banana,   encoder: mlp,     hidden: 256, mode: whiten,  lamb: 0.5}
+  banana_infonce:   {mixing: banana,   encoder: mlp,     hidden: 256, mode: infonce, sigma: 1.0}
+  sinusoid_lejepa:  {mixing: sinusoid, encoder: mlp,     hidden: 256, mode: lejepa,  lamb: 1.0e-3}
+  sinusoid_whiten:  {mixing: sinusoid, encoder: mlp,     hidden: 256, mode: whiten,  lamb: 0.5}
+  sinusoid_infonce: {mixing: sinusoid, encoder: mlp,     hidden: 256, mode: infonce, sigma: 1.0}
+  nvp_lejepa:       {mixing: nvp,      encoder: matched, n_layers: 8, mode: lejepa,  lamb: 1.0e-3}
+  nvp_whiten:       {mixing: nvp,      encoder: matched, n_layers: 8, mode: whiten,  lamb: 0.5}
+  nvp_infonce:      {mixing: nvp,      encoder: matched, n_layers: 8, mode: infonce, sigma: 1.0}
+
+alphas: [0.125, 0.25, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0]
+seeds: [1337, 1338, 1339]

JEPA/lejepa-identifiability/experiments/configs/grid.yaml

@@ -0,0 +1,27 @@
+# Grid search over lambda and rho (bound verification figure)
+# python run.py --config configs/grid.yaml --lamb 0.01 --rho 0.9 --seed 0
+
+experiment: grid
+out: results/grid
+
+# Data
+N: 2
+source_dist: gaussian
+num_eval: 10000
+
+# Training (shared)
+steps: 20000
+lr: 3.0e-3
+batch_size: 256
+log_every: 500
+
+# Encoder
+encoder: mlp
+hidden: 256
+mixing: spiral
+mode: lejepa
+
+# Sweep dimensions
+lambs: [1.0e-6, 1.0e-5, 1.0e-4, 1.0e-3, 5.0e-3, 1.0e-2, 5.0e-2, 1.0e-1, 5.0e-1]
+rhos: [0.3, 0.5, 0.7, 0.8, 0.9, 0.95, 0.99]
+seeds: [0, 1, 2]

JEPA/lejepa-identifiability/experiments/configs/reacher.yaml

@@ -0,0 +1,29 @@
+# Reacher pixel-observation identifiability experiment
+# Works for both OU and trajectory data — just point --data_dir at the right place.
+#
+# Prerender:
+#   python prerender.py eval
+#   python prerender.py ou --rho 0.95
+#   python prerender.py traj --delta 16 --h5_path data/reacher.h5
+#
+# Train:
+#   python run_reacher.py --config configs/reacher.yaml \
+#       --data_dir data/reacher/ou/rho=0.95
+
+experiment: reacher
+out: results/reacher
+data_root: data/reacher
+
+# Model
+d_latent: 2
+
+# Training
+epochs: 100
+batch_size: 256
+lr: 3.0e-3
+n_slices: 256
+n_eval_fast: 2000
+
+# Sweep dimensions
+lambs: [1.0e-3, 5.0e-3, 1.0e-2, 5.0e-2]
+seeds: [0, 1, 2]

JEPA/lejepa-identifiability/experiments/configs/scaling.yaml

@@ -0,0 +1,36 @@
+# Scaling experiment (paper figure)
+# python run.py --config configs/scaling.yaml --N 16 --seed 0
+# python run.py --config configs/scaling.yaml --N 16 --seed 0 --mode infonce
+# python run.py --config configs/scaling.yaml --N 16 --seed 0 --mode whiten
+
+experiment: scaling
+out: results/scaling
+
+# Data
+source_dist: gaussian
+num_eval: 10000
+
+# Training (shared)
+steps: 20000
+lr: 3.0e-3
+batch_size: 256
+rho: 0.95
+log_every: 500
+
+# Encoder
+encoder: matched
+n_layers: 4
+mode: lejepa     # default; override with --mode
+
+# Mode-specific defaults (used based on --mode)
+lamb: 1.0e-6     # for lejepa
+lamb_whiten: 0.5  # used when mode=whiten
+sigma: 1.0       # for infonce
+
+# Mixing
+mixing: coupling
+
+# Sweep dimensions
+dims: [2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+seeds: [0, 1, 2, 3, 4]
+K: 3  # parallel encoder runs per (N, seed); pick lowest loss