fix(phase-05/10): correct attention weights output + minor cleanups

- docs/en.md: printed weights were [0.446 0.312 0.242] but actual
  dot-product output on H + s_close_to_cat is [0.464, 0.305, 0.231].
  Fix to the true values so learners copy-pasting the example see
  matching numbers.
- code/main.py: drop unused 'context' variable from additive demo
  loop (CodeRabbit nit).
- outputs/prompt-attention-shapes.md: vary the three consecutive 'For'
  sentence openings for readability (CodeRabbit style nit).

Caught by CodeRabbit review of PR #48.

phases/05-nlp-foundations-to-advanced/10-attention-mechanism/code/main.py

@@ -67,7 +67,7 @@ def main():
     U_a = [[0.5, 0.2, 0.3], [0.2, 0.6, 0.2]]
     v_a = [0.8, 0.6]
     for name, s in [("close to 'cat'", [0.9, 0.1, 0.2]), ("close to 'mat'", [0.1, 0.9, 0.3])]:
-        context, weights = additive_attention(s, H, W_a, U_a, v_a)
+        _, weights = additive_attention(s, H, W_a, U_a, v_a)
         pretty = {p: round(w, 3) for p, w in zip(positions, weights)}
         print(f"  decoder state {name:20s} -> weights {pretty}")
 

phases/05-nlp-foundations-to-advanced/10-attention-mechanism/docs/en.md

@@ -117,7 +117,7 @@ print("weights:", w.round(3))
 ```
 
 ```
-weights: [0.446 0.312 0.242]
+weights: [0.464 0.305 0.231]
 ```
 
 First row wins. Then move the decoder state closer to the third encoder state and watch the weights shift. That is it. Attention is explicit alignment.

phases/05-nlp-foundations-to-advanced/10-attention-mechanism/outputs/prompt-attention-shapes.md

@@ -14,4 +14,4 @@ Given a broken attention implementation, you identify the shape mismatch. Output
 
 Refuse to recommend fixes that silently broadcast. Broadcast-hiding bugs surface later as silent accuracy degradation.
 
-For Bahdanau confusion, insist the decoder input is `s_{t-1}` (pre-step state). For Luong, `s_t` (post-step state). For dot-product, flag dimension mismatch between query and key as the most common first-time error.
+For Bahdanau confusion, insist the decoder input is `s_{t-1}` (pre-step state). For Luong, `s_t` (post-step state). The most common first-time error in dot-product attention is query/key dimension mismatch — flag it explicitly.