Make Week 2 verification boundaries explicit

[skyzh]

This is stacked on #157. That PR makes the Week 2 course order follow balanced benchmark evidence; this follow-up makes the limits of that evidence and the expected student workload explicit before Week 2 is treated as launch-ready.

The generic WIP banner is replaced with a verification matrix that separates Qwen3-0.6B macOS CI coverage, optional 1.7B/4B integration, single-machine M4 Pro performance evidence, and incomplete maintainer ownership. A checkpoint retention ledger records each optimization's invariant, hypothesis, winning and losing shapes, fallback, and benchmark failure mode. The course now distinguishes the Days 1–5 core path from the Days 6–7 performance lab, moves Xcode counter and Shader Cost Graph work into an optional advanced appendix, and identifies repeated dense-cache concatenation as an O(S²) copy baseline rather than a production cache design.

Decode attention now keeps explicit masks on the readable model path and validates the primitive over Qwen's D=128 boundary grid: query lengths 1 and 8, contexts around 32 and 128, GQA ratios 1 and 4, and causal and explicit masks. The operator benchmark exposes the same query-length, GQA-ratio, and mask controls so schedule efficiency can be measured instead of inferred from correctness.

Validation included the complete Week 2 reference suite (40 passed, 9 skipped), Ruff, an mdBook build, and Apple-GPU smoke benchmarks for causal L=8/GQA=4 and explicit-mask L=1/GQA=1 cases.

AI-assisted: Codex.