[kingdonb]

Summary

Comprehensive operational playbook and engineering feature story documenting the breakthrough performance milestones achieved on the Qwen3.6-35B-A3B Metal Engine on Apple Silicon M2 (24 GB Unified Memory) across Strata PRs #8, #9, #10, and #11.

Covered Breakthroughs

1. GQA Attention Fold (PR fix(frontend): remove hardcoded mock constants, static labels, and unreactive colors in dashboard #8): Eliminated 8x redundant KV cache memory traffic, preventing occupancy collapse and boosting deep-context decode by >4.5x.
2. Draft-Vocab Slicing (PR fix(frontend): Sovereign NOCK Stack chart clips data outside static 90k-115k YAxis domain #9): Reduced speculative decoding draft-head bandwidth by 30x–60x via in-place top-vocab projection slicing while maintaining 10/10 bit-identical text.
3. FP16 KV Cache (PR fix(worker): database snapshot deduplication overwrites generated_at and destroys historical time-series #10): Cut KV cache footprint exactly in half (0.34 GB at 16K ctx) with on-the-fly register unpacking.
4. Parallel Associative Prefix Scan (PR fix(worker): prevent silent mock fallback when primary RPC queries fail #11): Replaced token-sequential GDN recurrence with WY/UT associative prefix scan (14.6x faster kernel time).
5. 2D Tiled GEMM & Dequant-Once (PR fix(worker): prevent silent mock fallback when primary RPC queries fail #11): Eliminated redundant per-token dequantization with tiled matrix multiplication (15.3x faster kernel time).

Includes complete architecture diagram, empirical kernel microbenchmark tables, multi-turn telemetry logs (1,051 tokens @ 12.2–16.3 tok/s), strata-q36.json server configuration, and client querying recipes.