Aug 19

1. 21:54

   fix(retrieval & providers): cap max_context_chars to 3500 to keep prompt tokens < 1800 and update free model routing

   Mokhtar
   d1ecc9b
2. 16:06

   feat(production): complete phase 1-4 refactor with hybrid retrieval CAG architecture, E5 dense embeddings, grounding validator, precomputed cold-start index, and 24/24 passing system tests

   Mokhtar
   d54db1a