feat: EXP-15-19 training research, Gemma 4 adapter, data pipeline

[CalebisGross]

Summary

• EXP-15 through EXP-19 completed: rotation experiments, data scaling (12K encoding examples), Gemma 4 E2B integration
• Encoding spoke solved: 100% novel schema compliance on both Qwen 3.5 2B and Gemma 4 E2B
• Bespoke spoke models outperform Gemini 3 Flash on mnemonic's encoding task (100% vs 0% schema)
• Qwen 3.5 2B selected as production encoding model (1.7x faster than Gemma 4 locally at equal quality)
• Hallucination stress test: both spoke models 5/7, Gemini 1/7

Key additions

• gemma_spoke_adapter.py — Gemma 4 E2B spoke adapter (NF4, PLE CPU offload, SpokeWrappedLayer)
• batch_encode.py — Gemini Batch API pipeline for scalable training data generation
• compare_models.py — Side-by-side model comparison (schema, speed, quality)
• stress_test_hallucination.py — Hard input testing for detail preservation
• enrich_and_generate.py, extract_prenuke_data.py, merge_training_data.py — Data pipeline
• train_qwen_spokes.py — Updated with --model-type gemma support, OOM protection
• experiment_registry.md — Full results for EXP-15 through EXP-19
• Handoff encoding preserved verbatim in agent.go

Test plan

• Qwen novel eval: 10/10 (100%) schema compliance
• Gemma novel eval: 10/10 (100%) schema compliance
• Hallucination stress test: Qwen 5/7, Gemma 5/7
• Model comparison: Qwen 19.7s/input, Gemma 33.9s/input, Gemini fails
• End-to-end daemon integration via serve_spokes.py

🤖 Generated with Claude Code