Debug 503: Disable mount to isolate crash

Fix 503: Entrypoint Script + FUSE Config

Fix 503 Critical: Background mount + dos2unix + init

Fix 503: Run app even if mount fails

Fix 503: Fault-tolerant start (Run app even if mount fails)

Deploy Brain Bunker Bridge 2.0 (Clean)

feat: Implement Bunker Bridge 2.0 (Fail-Safe Azure Persistence)

v10.1.6: Restore missing PositionalEncoding class

v10.1.5: Add debug trace for VL-JEPA loading errors

v10.1.4: Restore missing constants Y_ENCODER_MODEL and FALLBACK_RERANKER

v10.1.3: Enable active training with AdamW optimizer

v10.1.2: Add einops dependency for nomic-embed

v10.1.1: Fix VL-JEPA train_step/analysis dimension logic

v10.1: Sequence-Based VL-JEPA (8 Layers, Node Tokens)

v10.0: VL-JEPA Paper-Faithful (Transformer 4L + nomic-embed)

v9.0: VL-JEPA Thought Predictor (L18.5) - arXiv:2512.10942

v8.0: Full Qwen3-VL Stack (L14 Embedding + L15 MRL + L16 Reranker)

v7.0: Qwen3-VL-Embedding-2B (2048D, MRL, Multimodal)

v6.0: Official Qwen3-Embedding-0.6B with native MRL support

v5.1: Upgrade to Qwen 3B (closest to 2B)

v5.0: Original Qwen 3-Layer Architecture (L14 Qwen + L15 MRL + L16 MiniLM)

v4.2: BGE-base-en 768D - proven to work on cpu-basic

v4.1: Fix - Use nomic-embed-text with MRL support instead of Qwen LLM

v4.0: Qwen L14-L15 Embedding + MiniLM L16 Reranker

Add sentence-transformers to requirements

v3.2: Switch to sentence-transformers (MiniLM) for reliable CPU embedding

v3.1: Lazy Loading - models load on-demand to save RAM

v3.0: Two-Stage RAG - Embedding + Reranker with /rerank endpoint

Enhanced multimodal API: EOS token, MRL (64-2048D), /embed_image, /embed_multi endpoints

Create Dockerfile

Create app.py

Create requirements.txt

initial commit