Summary

AI-assisted development diverges unless something deliberately counters it: similar problems yield similar-but-not-identical solutions, and diff-only review cannot see an already-existing helper. Cogitave's counter-force is retrieval before generation, stated as a hard rule in AI-Native Development (decision: ADR-0013): discover, compose, and only justify-and-author when nothing fits - then contribute the result back. That rule only works because there is exactly one canonical thing to reuse at each layer - tokens, primitives, golden components - queried through a single MCP-native registry, and it only stays real because anti-duplication gates (no-duplication CI, reuse review, AGENTS.md decision tables, and the eval harness's reuse metric) catch a needless re-author before it lands.

The same instinct governs whole engines and tools, not just components: ADR-0003 says port best-in-class references and rebuild them native only where the rebuild is materially better - faster, more integrated, more correct, more secure. Otherwise, the reference stays an explicit, time-boxed bootstrap crutch, never a hidden forever-dependency, and the Technology Radar records the default so any component's status - reference, bootstrap, or native primitive - is always answerable.

Building from scratch is not the default anywhere in this estate. It is a reasoned exception you can name, and now you can justify one - or recognize when you should not.