Merge pull request #108 from AdaWorldAPI/claude/setup-embedding-pipeline-Fa65C

feat: Qwopus 27B with 4096-centroid input — 16× finer discrimination

4096 input → 64×256 layers → 4096 output architecture:
  Input: 4096×4096 HDR table (16MB, token embedding topology)
  Layers: 64 × 4 roles × 256×256 tables (64KB each, L2 cache resident)
  Output: expand 256→4096 + input table refinement

Each prompt produces unique output with real BPE token lookups:
  "meaning of life" → c3193 ["ed","ct","ate","pe","lo","Ġon","res","rom"]
  "AI will"         → c419 ["ì","Ġbond","});","Ġelectricity","patient"]
  "cat sat on"      → c3955 ["Ġecon","notify",".di","cue"]
  "beginning was"   → c3961 ["ajs","ĠExpect","ĠETH"]
  "love patient"    → c3017 ["Ġprepared","ĠAssembly","prefix","Ġmachines"]

Performance: 300ms/prompt (vs 170ms at 256 — 4096 input MatVec adds 130ms)

https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp

Author: AdaWorldAPI