Aiki Alpha 3 reduces runtime overhead and adds adaptive number and list representations

Aiki Alpha 3 focuses on implementation efficiency without altering the language surface. The runtime now performs fewer allocations and traversals, reducing call overhead and literal realization costs. These changes preserve semantic invariants, meaning existing programs should behave identically but execute more efficiently. The improvements target core operations like argument frames and parser bookkeeping, which are common bottlenecks in dynamic languages.

Adaptive representations for numbers and lists introduce internal flexibility while maintaining exactness and immutability. Numbers can now use compact integers, rationals, or exact finite-binary forms as appropriate, and lists can switch between flat, frontier, or persistent forms. These choices remain hidden from programs, so no code changes are required. The trade-off is increased implementation complexity, but the payoff is reduced memory usage and faster operations for common cases.

Profiling and coverage tooling are now unified under a single observation network. This eliminates redundant runtime accounting mechanisms and provides a consistent view of calls, boundaries, allocations, and host costs. Coverage data now comes from the same execution evidence used for profiling, reducing overhead. The static coverable structure ensures that all program sites are accounted for, while dynamic hits provide runtime precision. This simplifies performance analysis but may require adjustments to existing profiling workflows.

The math surface is now explicitly separated by semantic role. Exact operations remain the default, while approximate math is confined to a dedicated module. This preserves Aiki's exact number model while still providing access to common approximate functions. The separation may require code changes for programs that previously relied on implicit approximate behavior. The FFI acceleration surface is also more constrained, reducing the risk of semantic drift between native and accelerated operations.

The repository structure is now more modular, with experiments moved to a separate project. This clarifies the boundary between language implementation and experimental work, reducing the risk of machine-specific code leaking into the core. The core repository is now more focused on the language itself, while experiments act as independent stress tests. Engineers contributing to Aiki will need to adapt to this new division, but the result should be a more maintainable and predictable codebase.