Add local Gemma-4-E4B (oMLX) evaluation harness for memory systems

[johnz1019]

Summary

Adds an evaluation harness to run the HaluMem memory-system benchmark fully locally against a Gemma-4-E4B model served via oMLX, with no external memory-system API keys required. The answer model, judge model, and embeddings are decoupled so each can be pointed at a local or self-hosted endpoint.

What changed

New local infrastructure

• local_mem0_components.py / local_embedding_server.py — deterministic blake2b hash embeddings + an OpenAI-compatible local embedding server, replacing OpenAI text-embedding-3-small.
• llms.py — separate answer vs judge model config; judge can read GPT-5.5 from Codex config (JUDGE_USE_CODEX_CONFIG) or JUDGE_OPENAI_* overrides; supports both Responses (streaming) and Chat wire APIs.

Local adapters for the memory systems

• eval_memzero_graph.py — run Mem0-Graph locally (local Neo4j + Qdrant + hash embeddings + local LLM).
• New: eval_gemma4_e4b_local.py (no-memory / full-session baseline), eval_memzero_local.py, eval_supermemory_local.py.
• Adapted: eval_memos.py, eval_memobase.py, eval_memzero.py, eval_supermemory.py, eval_zep.py for local endpoints.

Scoring & orchestration

• evaluation.py — route scoring through the new judge; add memory-extraction F1.
• run_gemma4_e4b_omlx_eval.py — single-system pipeline: start oMLX → run adapter → stop oMLX → judge scoring → write report.
• run_gemma4_e4b_memory_matrix.py — matrix runner across frames.
• write_gemma4_e4b_comparison_report.py — local vs README-baseline comparison report.

Notes

• Run outputs / data (eval/runs/, eval/reports/, .supermemory/) are intentionally not included.
• No secrets are committed (.env is gitignored; verified the diff contains no keys).
• Implementation authored by codex on the lay2-studio host; packaged and submitted here.