Performance measurements

Run the production-artifact benchmark with:

make bench

The benchmark uses one scope-precedence request for TypeScript, Go, Rust and PHP in both ast and generated modes. Every row receives the same assign and define values, renders 177 UTF-8 bytes with the same SHA-256, and runs with the final-page cache disabled. A mismatched output, repeated output or prepared output aborts the run before results are written.

Each row contains 21 independent cold-process samples and 21 independent persistent-process samples of 1,000 renders after 20 warmups. The report records median and P95 for:

• compiler time: source parsing and AST artifact creation; generated mode also includes typed IR lowering and host-source emission;
• cold process time and peak RSS;
• full Program.render time, including request binding;
• prepared render time after one Program.prepare call;
• persistent process time per render and peak RSS;
• artifact bytes and exact output identity.

PreparedRender is a separate metric because it measures generated control flow without repeatedly normalizing the same request. It is the relevant path when one normalized request is rendered more than once. Full render remains the request-level cost and must be considered for one-shot rendering.

The current 21-sample run produced these medians:

Language	Mode	Cold process	Full render	Prepared render	Persistent RSS
TypeScript	AST	137.41 ms	0.0080 ms	0.0020 ms	94.25 MiB
TypeScript	generated	131.65 ms	0.0073 ms	0.0010 ms	92.97 MiB
Go	AST	7.11 ms	0.0030 ms	0.0016 ms	11.34 MiB
Go	generated	8.60 ms	0.0038 ms	0.0008 ms	10.70 MiB
Rust	AST	9.64 ms	0.0051 ms	0.0024 ms	2.80 MiB
Rust	generated	5.81 ms	0.0043 ms	0.0017 ms	2.53 MiB
Php	AST	70.12 ms	0.0178 ms	0.0106 ms	28.08 MiB
Php	generated	81.61 ms	0.0130 ms	0.0045 ms	28.22 MiB

Generated mode improves prepared rendering in all four implementations. Go and Rust show that request binding can dominate a very small page: Go generated full render is slower than AST even though its prepared renderer is faster, while Rust is approximately equal at full-render scale. Keep both measurements when choosing a mode; a prepared-render number alone does not describe complete request cost.

The committed JSON contains the exact median, P95, compiler and artifact-size values used by the example site. Absolute values depend on the machine and toolchain, so compare modes from one run rather than combining results from different runs.