Design decisions

Not manuals — records of why it was done this way and not another way. For users who want to trust the library's judgment, and contributors who would otherwise re-litigate (or silently break) a settled trade-off.

• cache.md — why caching defaults on, and why the global override beats explicit per-call arguments
• collections.md — why jitclass, why fixed dtypes, why counter is a function
• rng.md — why our own seed(), why the alias method is a split API
• parallel.md — why complete operations instead of prange wrappers, and every threshold in them
• benchmarking.md — why the tooling measures the way it does, and what CI deliberately doesn't report

Each document states the decision, the alternatives considered, and the reason the alternative lost. If a future change invalidates a reason, the decision is open again — that's what these files are for.