Performance, with the gaps visible.
Start with the slowdowns that deserve investigation, inspect each old/current comparison, then check whether the evidence is strong enough to support a release claim.
The charts use the same accepted structured results as the contextual cards throughout this guide. The latest Android data is the July 27 packaged migration recheck: all 11 matched medians and p95s beat the frozen released implementation, while a second production-build device and calibrated runs remain open.
The accepted Swift/macOS measurements remain visible as a historical audit, but new Swift performance runs are temporarily deferred with the other Apple-only verification gates. They are not current release evidence.
Largest measured gaps
Positive means slower than the frozen implementation. Only deltas above +5% appear here, so small runner noise does not dominate the queue.
Kotlin results are advisory until the JMH gate exists. Swift event parsing remains inside the descriptive 10% median budget. Use this view to choose the next investigation, not to compare hardware.
Relative median delta
Each row compares old and current on the same runner or device. Negative values are improvements; positive values identify a gap.
p95 amplification
The widest current p95-to-median ratios for Android. Larger bars point to scheduling, warm-up, allocation, or thermal variance worth isolating.
What still blocks stronger claims
A fast median does not close platform coverage, repeatability, or resource-cost gaps.
Calibrate five unchanged-revision runs on a stable host before enforcing latency gates.
2026-07-26 · 5e7c0c65e6e8 · full auditReplace the plain loop with JMH forks, warm-up, Blackhole, and allocation profiling.
2026-07-26 · 5e7c0c65e6e8 · full auditReproduce on a second production-build device, then calibrate unchanged-revision runs.
2026-07-27 · 27e135eac626 · full auditRun the packaged facade on physical iOS and capture startup, memory, archive size, and dSYMs.
2026-07-26 · 5e7c0c65e6e8 · full auditRun real browsers and Workers, including transfer and startup measurements.
2026-07-26 · 5e7c0c65e6e8 · full auditWhat is included
The harnesses call authored public APIs, so measurements include language adaptation and FFI or Wasm overhead. Host comparisons use three paired old/current processes. Android uses three non-debuggable AndroidX Benchmark invocations with 30 measurements each and adjacent legacy/SDK pairs.
Negative deltas mean the current SDK is faster. Cached accessors below the clock's useful floor are API-shape checks and are excluded from gap charts. Current-only operations establish a descriptive baseline but have no old-library delta.
Audit records
The interactive view is for analysis; the full reports preserve exact scope, environment, methodology, decisions, and developer guidance as regular, navigable website pages.
Physical Pixel results, allocation isolation, and the remaining second-device gate.
Open report RustShared-core auditNative latency, artifact footprint, retained optimizations, and release guidance.
Open report Kotlin/JVMJVM facade auditPublic-facade crossings, current limitations, and the work required for a JMH gate.
Open report TypeScriptWebAssembly auditNode and Wasm facade measurements, package footprint, and browser evidence gaps.
Open report SwiftHistorical Apple auditAccepted macOS measurements retained for context while new Apple verification is deferred.
Open report MethodologyMeasurement contractExact revisions, comparison rules, regression budgets, footprint policy, and open gates.
Open reportThese results are descriptive for their recorded runner or device. They are not cross-hardware rankings or universal application-latency guarantees.