Fix video caption overlap with speed controls, merge Build on this section

app/src/components/Video.astro

@@ -96,6 +96,7 @@ const id = `video-${Math.random().toString(36).slice(2, 9)}`;
     align-items: center;
     gap: 6px;
     margin-top: 8px;
+    margin-bottom: 4px;
   }
   .speed-label {
     font-size: 0.8rem;

app/src/content/chapters/folding/01-hero.mdx

@@ -11,7 +11,7 @@ To change this, we trained a robot on a challenging but highly requested task: *
 
 <Video src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/level2.mp4" />
 
-<p style="text-align: center; color: var(--muted-color); font-size: 0.85rem; margin-top: -8px;">Autonomous folding of crumpled t-shirts (Level 2)</p>
+<p style="text-align: center; color: var(--muted-color); font-size: 0.85rem; margin-top: 0;">Autonomous folding of crumpled t-shirts (Level 2)</p>
 
 --------
 

app/src/content/chapters/folding/08-ablations.mdx

@@ -164,7 +164,7 @@ Here is an uncut Level 1 evaluation run from Experiment 2.5, 15 minutes of conti
 
 <Video src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/level1.mp4" />
 
-<p style="text-align: center; color: var(--muted-color); font-size: 0.85rem; margin-top: -8px;">Autonomous folding from flat state (Level 1)</p>
+<p style="text-align: center; color: var(--muted-color); font-size: 0.85rem; margin-top: 0;">Autonomous folding from flat state (Level 1)</p>
 
 -------------
 

app/src/content/chapters/folding/09-learnings.mdx

@@ -60,6 +60,4 @@ lerobot-train \
 
 End to end, this project took roughly ~1,900 engineering hours: designing and building setups, collecting data, implementing new features in LeRobot, running training, evaluating on hardware, designing the OpenArm Mini teleop system, and feeding everything back upstream. Across 11 experiments, we varied models, action representations, and reward weighting. Algorithmic changes moved success rates by 5–20 percentage points. Finetuning on a curated dataset one-fifth the size moved it by 50. The bottleneck wasn't the model, it was the data. The single biggest lever was **clean, consistent data** and the tooling to score and curate it, which now exists in LeRobot.
 
-### Build on this
-
 We open-sourced everything: code, data, models, reward model, and now the full story of what worked and what didn't, so the community can push further. Use our [dataset](https://huggingface.co/datasets/lerobot/high_quality_folding) directly, try new architectures, experiment with different recipes. If you find something that works, we'd be happy to run your models on our physical setups and share the results back.