Fix bugs, N+1 queries, and wire settings in persian-tutor

- Replace inline __import__("datetime").timedelta hack with proper import
- Remove unused import random in anki_export.py
- Add error handling for Claude CLI subprocess failures in ai.py
- Fix hardcoded absolute path in stt.py with relative Path resolution
- Fix N+1 DB queries in vocab.get_flashcard_batch and dashboard.get_category_breakdown
  by adding db.get_all_word_progress() batch query
- Wire Ollama model and Whisper size settings to actually update config
  via ai.set_ollama_model() and stt.set_whisper_size()

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

python/persian-tutor/stt.py

@@ -1,13 +1,17 @@
 """Persian speech-to-text wrapper using sttlib."""
 
 import sys
+from pathlib import Path
 
 import numpy as np
 
-sys.path.insert(0, "/home/ys/family-repo/Code/python/tool-speechtotext")
+# sttlib lives in sibling project tool-speechtotext
+_sttlib_path = str(Path(__file__).resolve().parent.parent / "tool-speechtotext")
+sys.path.insert(0, _sttlib_path)
 from sttlib import load_whisper_model, transcribe, is_hallucination
 
 _model = None
+_whisper_size = "medium"
 
 # Common Whisper hallucinations in Persian/silence
 PERSIAN_HALLUCINATIONS = [
@@ -18,11 +22,19 @@ PERSIAN_HALLUCINATIONS = [
 ]
 
 
-def get_model(size="medium"):
+def set_whisper_size(size):
+    """Change the Whisper model size. Reloads on next transcription."""
+    global _whisper_size, _model
+    if size != _whisper_size:
+        _whisper_size = size
+        _model = None
+
+
+def get_model():
     """Load Whisper model (cached singleton)."""
     global _model
     if _model is None:
-        _model = load_whisper_model(size)
+        _model = load_whisper_model(_whisper_size)
     return _model