Add a benchmark for memory usage

[fabian-hiller]

Here some X post about it:

• https://x.com/bentlegen/status/2084750022063579522
• https://x.com/irvinebroque/status/2084084405803655511

Would be great to have a prof of it

[EskiMojo14]

per brief copilot discussion (putting here so i can remind myself later):

What's actually needed

1. Run Node with --expose-gc — add the flag in orchestrate.ts:

execFile(
  process.execPath,
  ["--expose-gc", path.resolve(..., "library.ts"), `--lib=${lib}`],
  ...
)

2. Schema footprint — measure heap delta around initialization runs:

global.gc?.();
const before = process.memoryUsage().heapUsed;
const result = run();
global.gc?.();
const heapBytes = process.memoryUsage().heapUsed - before;

This works cleanly because the process only ever holds one library's allocations.

3. Per-call allocation (GC pressure) — run a secondary tight loop before forcing GC to capture bytes allocated rather than bytes retained:

const SAMPLES = 10_000;
const before = process.memoryUsage().heapUsed;
for (let i = 0; i < SAMPLES; i++) run(data);
const allocatedPerCall = (process.memoryUsage().heapUsed - before) / SAMPLES;

heapUsed after GC would always be ~0 for pure validation functions since they don't retain anything — you want the pre-GC snapshot to measure allocation rate.

Remaining considerations

• heapUsed is still noisy at small scales; you'd want to run this after tinybench's warmup phase so JIT is stable
• heapBytes on the schema footprint would be a new optional field on baseBenchResultSchema alongside mean
• The initialization result type is the natural home for schema footprint; validation/parsing results would carry the allocation-rate metric