Day 3: correct gitignore, add data manifest

.gitignore

@@ -1,6 +1,7 @@
 venv/
 data/raw/
 data/processed/*.csv
+data/indexes/
 models/
 *.gguf
 __pycache__/
@@ -10,4 +11,3 @@ __pycache__/
 dist/
 .DS_Store
 .claude/
-claude_code_overnight_prompt.md

data/MANIFEST.md

@@ -0,0 +1,22 @@
+# Data Manifest
+Files tracked here are too large for git. Re-download using instructions below.
+
+## data/raw/ — source datasets (not in git)
+| Dataset | Source | Size | Download command |
+|---|---|---|---|
+| GoEmotions | HuggingFace | ~200MB | python src/data/download_datasets.py |
+| Empathetic Dialogues | HuggingFace | ~500MB | python src/data/download_datasets.py |
+| Reddit Mental Health | zenodo.org/records/3941387 | 3.1GB | zenodo_get 3941387 -o data/raw/reddit_mental_health/ |
+| Suicide Detection | kaggle nikhileswarkomati/suicide-watch | 60MB | kaggle datasets download -d nikhileswarkomati/suicide-watch -p data/raw/suicide_detection/ --unzip |
+
+## data/indexes/ — built artifacts (not in git)
+| File | How to rebuild |
+|---|---|
+| faiss_flat.index | Run 
+otebooks/colab_build_faiss_index.ipynb on Colab A100 |
+| metadata.db | Same as above |
+
+## data/processed/ — derived files (not in git)
+| File | How to rebuild |
+|---|---|
+| nli_train/val/test.csv | python src/data/build_nli_pairs.py |