Block 3: Calendar Optimization & Phase 4 Lessons

Block 3: Calendar Optimization & Phase 4 Lessons

8 lessons from Phase 3 completion and Phase 4 (calendar optimization) implementation.

# Lesson Core Teaching
020 DuckDB executemany Hangs — Use PyArrow executemany hangs at 12K+ rows; PyArrow bulk insert completes in <1s
021 Calendar as Portfolio Optimization Selecting a collection is constrained set optimization, not top-N ranking
022 Heuristic Month-Fit Without Text Visual features alone provide coarse but sufficient seasonal signal
023 Maximum Marginal Relevance Greedy MMR with CLIP cosine similarity balances quality and diversity in <1s
024 Hungarian Algorithm for Assignment scipy's linear_sum_assignment gives provably optimal 1-to-1 assignment
025 Multiple Methods as Baselines Naive baseline + alternatives prove the optimizer adds value
026 Formalizing De Facto Dependencies Imported but undeclared dependencies break on fresh installs
027 Migration Ordering and Apply-on-Use Migration on disk ≠ migration applied; test the full write path