docs: localize images to docs/images for Sphinx static building

Author: Vishal-sys-code

docs/architecture.md

@@ -4,7 +4,7 @@ The **Spiking Decision Transformer (SNN-DT)** is a novel integration of biologic
 
 SNN-DT addresses this fundamental limitation by introducing three foundational neuromorphic innovations into the return-conditioned Transformer pipeline.
 
-![SNN-DT Architecture Overview](../model_architecture.png)
+![SNN-DT Architecture Overview](images/model_architecture.png)
 
 *Figure 1: Overall architecture of the Spiking Decision Transformer. The offline trajectory embeddings (return, state, action) pass into the Phase-Shifted Positional Encoder. Spiking Self-Attention blocks interact dynamically with a Dendritic Routing MLP, culminating in a sparse action projection governed by Three-Factor Plasticity.*
 
@@ -22,7 +22,7 @@ This mechanism endows each attention head with a distinct rhythmic code, providi
 
 SNN-DT maps continuous embeddings into sparse spike rates via Leaky Integrate-and-Fire (LIF) dynamics. To adaptively combine parallel attention head outputs without uniform averaging, we leverage a lightweight **Dendritic-Style Routing MLP**.
 
-![Dendritic Routing Coefficients](../visualizations/routing_coefficients_heatmap.png)
+![Dendritic Routing Coefficients](images/routing_coefficients_heatmap.png)
 
 *Figure 2: Heatmap illustrating dynamic, context-dependent routing gating coefficients $\alpha^{(h)}_i(t)$ per head across a sequence. Darker cells indicate effectively pruned or suppressed head outputs, securing computational sparsity.*
 

docs/experiments.md

@@ -8,7 +8,7 @@ Our evaluation specifically tracks (1) algorithmic performance and (2) proxy met
 
 We isolate the impact of Phase-Shifted Positional Spiking (Pos-Only) and Dendritic-Style Routing MLP (Route-Only) against a unified configuration (Full) and the base non-augmented LIF formulation (Baseline).
 
-![Offline Loss Validation Curves](../visualizations/offline_loss_curves.png)
+![Offline Loss Validation Curves](images/offline_loss_curves.png)
 
 *Figure 1: Ablation validation loss trajectories. The Full model natively achieves the fastest convergence towards the error floor by exploiting highly diverse temporal encoding and responsive gating.*
 
@@ -19,7 +19,7 @@ We isolate the impact of Phase-Shifted Positional Spiking (Pos-Only) and Dendrit
 | **Acrobot-v1** | $-87.1 \pm 3.2$ | $-72.0 \pm 3.6$ | $-68.3 \pm 3.9$ | $\mathbf{-59.7 \pm 2.7}$ |
 | **Pendulum-v1** | $-155.3 \pm 5.1$ | $-140.0 \pm 4.7$ | $-135.4 \pm 4.4$ | $\mathbf{-130.5 \pm 4.2}$ |
 
-![RL Performance plot](../visualizations/rl_performance_plot.png)
+![RL Performance plot](images/rl_performance_plot.png)
 
 *Figure 2: Performance distributions evaluated over the target environments tracking downstream RL validation. The density directly reflects tighter policy resilience in continuous evaluations.*
 
@@ -29,7 +29,7 @@ We isolate the impact of Phase-Shifted Positional Spiking (Pos-Only) and Dendrit
 
 On advanced neuromorphic substrates like Intel Loihi or IBM TrueNorth, algorithmic energy scales linearly with spike activity emissions. We compute absolute spike counts during test batches as an energy proxy.
 
-![Spike Emission Distribution](../visualizations/spike_histogram.png)
+![Spike Emission Distribution](images/spike_histogram.png)
 
 *Figure 3: Histograms of localized sparse spike activity. SNN-DT networks suppress superfluous event spikes effectively limiting output variance beneath the 10-spike barrier compared to unrestricted formulations.*