Fix Jina Embedder Issues and Config Integration

[r3d91ll]

Jina Embedder Critical Issues

Found during code review. These issues affect performance, reliability, and configuration management.

🔴 CRITICAL Issues (Performance/Functionality)

1. Redundant Encoding Bug - Line 620-771

   • Problem: _chunk_with_context_windows encodes document N times for N chunks
   • Impact: O(N²) complexity instead of O(N), massive performance degradation
   • Fix: Encode once, then chunk the embeddings

2. Config Extraction Broken - Lines 79-85

   • Problem: When EmbeddingConfig passed, only extracts device/batch_size/use_fp16
   • Missing: chunk_size_tokens, chunk_overlap_tokens, model_name (though model_name standardization is intentional)
   • Fix: Extract all config fields, especially chunking parameters

3. Hardcoded Chunk Parameters - Lines 83-84

   • Problem: Forces chunk_size_tokens=1000, chunk_overlap_tokens=200
   • Impact: Can't tune chunking for different document types
   • Fix: Use config values if provided

🟡 HIGH Priority Issues

4. No CUDA Check - Line 98

   • Problem: Defaults to 'cuda' without checking torch.cuda.is_available()
   • Impact: Crashes on CPU-only systems
   • Fix: Check CUDA availability, fallback to CPU

5. Missing Multimodal Guard - Line 359

   • Problem: Calls model.encode_image without checking if method exists
   • Impact: AttributeError on text-only models
   • Fix: Add hasattr check or disable multimodal for text-only variants

🟢 MEDIUM Priority Issues

6. No Truncation Warning - Line 800

   • Problem: Silently truncates documents > 32k tokens
   • Fix: Log warning when truncation occurs

7. Missing Type Annotations

   • Problem: No type hints on most methods
   • Impact: Fails mypy strict checks
   • Fix: Add comprehensive type annotations

Architecture Context

• Standardizing on Jina v4 is intentional (one model, one RAG architecture)
• 32k context window enables full paper processing
• Late chunking preserves context per HADES framework

Config Integration Needed

• Move chunk parameters to YAML config in /core/config/workflows/
• Create embedder-specific config files
• Ensure CLI args properly reach workers

Related Issues

• Critical: Fix production issues in workflow_arxiv_sorted_simple.py #44 - Workflow critical fixes (storage drain, GPU assignment, etc.)
• Feature/dual embedders late chunking 42 #43 - Chunking edge case fixes

Test Requirements

• Test redundant encoding fix with various document sizes
• Verify config parameter passing
• Test CPU fallback behavior
• Benchmark performance improvements