Make Week 2 follow measured bottlenecks

[skyzh]

Week 2 is meant to teach an evidence loop, but the chapters need more than an end-to-end number to justify each next step. The earlier operator runner measured implementations in fixed blocks, so GPU performance-state drift could consistently favor whichever implementation ran later. It also hid two dispatch decisions that did not survive a matched operator comparison.

This balances every readable, course, and MLX operator order and rotates kernel-group order before rerunning the complete progression. Each chapter from Day 2 onward now ends with the model throughput, operator latency, attribution profile, and selective Xcode evidence needed to choose the next bottleneck. Day 2 also teaches students how to capture the same counter tables and Shader Cost Graph for their own implementation.

The paired evidence changes two retained choices. Fused decode attention wins through context 128 but loses beyond that measured crossover, so the dispatch guard follows the data. Direct quantized embedding is slower than the readable packed row lookup at this checkpoint, so Week 2 drops it while Week 3 retains and tests it where the serving path uses the fused operator.

On the balanced M4 Pro run, the final 128-token Week 2 checkpoint reaches 95.4% of MLX prefill, 84.4% of MLX decode, and 85.3% of MLX output throughput. Split-K improves complete 32-token prefill by 11.9% and remains neutral at 128 tokens. The appendix attaches those fresh-process samples beside projection microbenchmarks, dependency-aware attribution, dispatch geometry, and the source-enabled Xcode matvec profile.

AI-assisted: Codex.