JS bench honesty audit

Same methodology as the Rust audit (rust-honest-disassembly.md): run each suspect workload, capture BENCH_SINK (the +=-accumulator of every iteration's workload result), divide by the expected single-call result, and check that the quotient matches the iteration count the harness believed it did.

If the wrapper says "I ran N iters in T seconds" and the sink equals "N × workload_result", every iteration actually executed. If the sink equals "1 × workload_result", V8 collapsed the outer loop the way HotSpot had done.

Results

All six quotients exactly match the iteration count the adaptive loop expanded to. V8 / TurboFan is NOT folding the outer timing loop on any of these workloads. The reported JS times are real per-iteration costs.

Why V8 doesn't fold

Dynamic typing. V8's optimiser has to assume workload() could return any JavaScript value (the type-feedback array only profiles the first ~100 calls). The += accumulator runs through the _arith('+', ...) runtime helper to handle string/number mixing, which TurboFan cannot prove pure. So even when the workload is constant-foldable in principle ((1 to 10).map.filter.fold = 36), V8 still emits the loop and evaluates each iter.

The result is that the JS wrapper measures correctly without needing the JMH-style AtomicLong.getAndAdd anti-fold that the JVM wrapper required.

Action

None. JS numbers in the bench table are real measurements. The remaining work on JS is genuine perf (the _arith indirection and _dispatch method routing dominate hot loops), not bench infrastructure.