Measured, not promised
Every library's benchmarks are flattering. Not because authors lie — because benchmarks lie by default, and it takes real work to stop them. Building ivue's numbers, our own harnesses lied to us three times. This post is the confession log, because the failure modes live in your benchmarks too.
Lie one: the JIT deleted the competition. Our first creation benchmark wrote instances into one overwritten variable. V8's escape analysis elided the competitors' allocations entirely — reactive() "created" 100,000 proxies in 0.0 ms — while ivue's opaque construction kept honestly paying. The fix: every instance retained in an array, touched after timing. Nothing the optimizer can discard.
Lie two: the unobserved heap. Our first memory benchmark read instances outside any effect — and reactive() looked impossibly light, because a proxy allocates its dependency storage only when something subscribes. Live objects are observed objects. Re-measured with every instance read inside its own subscribing effect, the proxy's real bill arrived: dependency maps for every tracked key.
Lie three: the harness tax. Timing operations through a per-call closure added a constant cost to every variant — large enough to hide a 3× difference between two method-call styles. Loop inside the timed region, or measure the harness instead of the code.
What survived the discipline ships on the Performance by Design page with its protocol attached — and the sharpest habit we kept: run the numbers in the reader's browser, on the shipped engine, where no one can quietly pick the machine:
If a claim can't survive its own methodology section, it isn't a result — it's marketing with axes. Steal the harness patterns; they're described under Methodology.