feat: ModernBERT architecture incompatibility chain fix and gitignore cleanup

Dockerfile

@@ -30,15 +30,20 @@ RUN pip install --no-cache-dir \
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# ── Layer 4: Pre-cache ALL ML models at build time ───────────
-# This eliminates cold-start latency on first request.
-# Models are cached to /app/models (persisted across container restarts).
+# ── Layer 4a: Pre-cache Nomic Embedding model ────────────────
+# Separated from reranker to prevent one failure from blocking the other.
 RUN python -c "\
-from sentence_transformers import SentenceTransformer, CrossEncoder; \
+from sentence_transformers import SentenceTransformer; \
 SentenceTransformer('nomic-ai/nomic-embed-text-v1', cache_folder='/app/models', trust_remote_code=True); \
-print('✓ Nomic Embedding model (768-dim) cached'); \
-CrossEncoder('cross-encoder/ettin-reranker-68m-v1', max_length=512); \
-print('✓ Ettin Reranker cached')"
+print('✓ Nomic Embedding model (768-dim) cached')"
+
+# ── Layer 4b: Pre-cache Ettin Reranker (ModernBERT) ──────────
+# CRITICAL FIX: trust_remote_code=True is MANDATORY for ModernBERT architecture.
+# Without it: KeyError: 'modernbert' → build crash.
+RUN python -c "\
+from sentence_transformers import CrossEncoder; \
+CrossEncoder('cross-encoder/ettin-reranker-68m-v1', max_length=512, trust_remote_code=True); \
+print('✓ Ettin Reranker (ModernBERT) cached')"
 
 # ── Layer 5a: Pre-cache IndicTrans2 English→Indic model ──────
 # This is the English→Indic direction model (Sprint 18).

api.py

@@ -115,7 +115,7 @@ def get_reranker_model():
     global reranker_model
     if reranker_model is None:
         print("⏳ Loading Cross-Encoder Reranker Model...")
-        reranker_model = CrossEncoder('cross-encoder/ettin-reranker-68m-v1', max_length=512, device='cpu')
+        reranker_model = CrossEncoder('cross-encoder/ettin-reranker-68m-v1', max_length=512, device='cpu', trust_remote_code=True)
         print("✅ Reranker Model loaded")
     return reranker_model
 

requirements.txt

@@ -11,7 +11,7 @@ uvicorn[standard]
 # --- ML / AI (CPU-ONLY WHEELS) ---
 # CRITICAL: --extra-index-url in Dockerfile forces CPU torch
 # DO NOT add torch here — it is installed separately in Dockerfile Layer 2
-sentence-transformers>=2.7.0
+sentence-transformers>=4.0.0
 einops
 huggingface-hub