Implement numerical features and episodic memory

[tylerbessire]

Summary

• Extend feature extraction to handle raw grids safely and expose a deterministic numerical feature vector
• Add a fully-fledged episodic memory with serialisable Episode objects and similarity queries
• Harden neural guidance and solver fallbacks for reliable operation

Testing

• pytest tests/test_features.py tests/test_memory.py tests/test_dsl_ops.py tests/test_dsl_cache.py -q
• pytest -q

---

https://chatgpt.com/codex/tasks/task_e_68c1dfadb4908322bca8bbca6eab0e91

Summary by CodeRabbit

• New Features

  • Added numeric feature vector output for training pairs.
  • Introduced persistent episodic memory with save/load, similarity search, and richer episode metadata.
  • Solver now returns two attempts for each test input.

• Bug Fixes

  • Improved handling of single-sample predictions; outputs are correctly shaped 1D probabilities.
  • Safer transformation hint detection and analysis when inputs are empty or malformed.
  • Robust parsing of training/test grids; invalid items are skipped or gracefully defaulted.

• Refactor

  • Split feature computation into original vs. canonical representations for more reliable stats and hints.

[coderabbitai]

Caution

Review failed

The pull request is closed.

Walkthrough

Introduces a two-stage feature pipeline and numeric feature vectorization; adjusts classifier forward/prediction shapes; redesigns episodic memory with persistent episodes, indices, and similarity retrieval; and hardens solver control flow with enhanced/baseline fallbacks, malformed data guards, and dual attempt outputs.

Changes

Cohort / File(s)	Summary of changes
Feature extraction pipeline arc_solver/features.py	Adds original_pairs → canonical_pairs flow, shifts certain stats to original_pairs and others to canonical_pairs, updates hint detectors with empty-input guards, changes object count preservation semantics, adds compute_numerical_features(train_pairs) returning deterministic numeric vectors, broadens error handling.
Classifier output shaping arc_solver/guidance.py	Reshapes 1D inputs to 2D in forward, squeezes single-sample outputs to 1D (length 7), ensures predict_operations returns 1D probabilities via ravel(), and casts to float for threshold comparisons.
Episodic memory persistence and API arc_solver/memory.py	Introduces Program alias, Episode dataclass with auto feature computation and (de)serialization, EpisodeDatabase with persistent storage, indices, similarity retrieval, and expanded API; refactors EpisodicRetrieval to use the new database, caching, and standardized methods.
Solver robustness and dual attempts arc_solver/solver.py	Adds enhanced→baseline fallback when degenerate zero outputs detected, skips malformed train/test grids via try/except with placeholders, handles no-training-data by echoing inputs for both attempts, enhances baseline synthesis fallbacks, and returns both attempt_1 and attempt_2 outputs.

Sequence Diagram(s)

sequenceDiagram
  participant C as Caller
  participant S as Solver
  participant E as EnhancedSolver
  participant B as BaselineSolver

  C->>S: solve_task(train_pairs, test_inputs)
  S->>S: sanitize/convert train & test (skip bad, use placeholders)
  alt have valid training data
    S->>E: run enhanced attempt_1
    alt enhanced ok and non-degenerate
      E-->>S: attempt_1
    else enhanced fails or degenerate
      E-->>S: error/zero-grid
      S->>B: fallback to baseline
      B-->>S: attempt_1 (with sanity fallback)
    end
    S->>B: generate attempt_2 (if available)
    B-->>S: attempt_2 or fallback to echo
  else no valid training data
    S-->>C: return [echo inputs, echo inputs]
  end
  S-->>C: return [attempt_1, attempt_2]

Loading

sequenceDiagram
  participant U as User
  participant ER as EpisodicRetrieval
  participant DB as EpisodeDatabase

  rect rgba(200,230,255,0.25)
  note right of ER: Store successful solution
  U->>ER: add_successful_solution(train_pairs, programs, task_id)
  ER->>DB: store_episode(signature, programs, task_id, train_pairs, metadata)
  DB->>DB: compute features, index by signature/program
  DB-->>ER: episode_id
  end

  rect rgba(220,255,220,0.25)
  note right of ER: Query by similarity
  U->>ER: query_for_programs(train_pairs, max_candidates)
  ER->>DB: query_by_similarity(train_pairs)
  DB->>DB: compute features, similarity scores
  DB-->>ER: candidate episodes/programs
  ER-->>U: candidate programs
  end

Loading

sequenceDiagram
  participant FE as FeatureExtractor
  participant OP as original_pairs
  participant CP as canonical_pairs

  FE->>OP: cast grids to int
  FE->>FE: compute basic stats from OP
  OP-->>CP: canonicalize_D4
  FE->>CP: analyze colors/bg/object counts
  FE->>FE: detectors (rot/ref/translate/etc.) with empty-guards
  FE-->>FE: compute_numerical_features() -> fixed-order numeric vector

Loading

Estimated code review effort

🎯 5 (Critical) | ⏱️ ~120 minutes

Poem

I nudge the grids, then twirl them twice,
Pack memories snug like carrots on ice. 🥕
Classifier hums a seven-note tune,
Solver tries two paths under the moon.
Canon to original, features align—
Hop, store, retrieve—solutions shine!

---

📜 Recent review details

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between ebbecb0 and 7e33527.

📒 Files selected for processing (4)

• arc_solver/features.py (8 hunks)
• arc_solver/guidance.py (1 hunks)
• arc_solver/memory.py (1 hunks)
• arc_solver/solver.py (1 hunks)

✨ Finishing Touches

• 📝 Generate Docstrings

🧪 Generate unit tests

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch codex/continue-implementation-for-sota-model

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}