ZEngine

Status: CPU instrumentation facilities, arena tracking, and render-graph timestamp collection exist. Frame integration, presentation, and validation coverage are incomplete.

Current facilities

Profiling.h selects one of three compile-time paths:

Condition	Scope/value/alloc behavior
ZENGINE_PROFILING && ZENGINE_TRACY	Tracy zones, plots, and allocations
ZENGINE_PROFILING	ScopeGuard/ProfilerBuffer fallback and MemoryProfiler alloc hooks
neither	profiling macros expand to nothing

ZENGINE_PROFILING is currently supplied for Debug and RelWithDebInfo through the engine CMake target. ZENGINE_TRACY is an optional root CMake option that defaults to OFF; it is not implicitly enabled for Debug.

ProfilerBuffer is a fixed-capacity (4096) static CPU sample/value buffer. It permits concurrent writers and exposes the previous-frame snapshot through DumpLastFrame and DumpLastFrameValues. Samples are silently dropped at capacity. Its BeginFrame() transition and its dump APIs are facilities, not evidence that the main loop currently calls or presents them.

MemoryProfiler records named arenas registered through MemoryManager::CreateBudgetedArena, tracks current/peak offsets, emits profile values, and warns above 80% with a 60-second cooldown. It is initialized with the Bootstrap owner before tracking. Engine::MainThreadRun() calls MemoryProfiler::Update() at its main-thread frame boundary, so the editor panel reflects live named-owner current and peak data. The panel publishes separate render-thread snapshots for VMA/driver heap samples, persistent environment resources, and render-graph transients; it intentionally does not collapse these into a CPU-arena total.

RenderGraph owns per-frame Vulkan timestamp query pools when supported by the selected queue. It records RGPassTiming values and exposes GetLatestPassTimings() / IsTimestampProfilingEnabled(). Query-pool creation failure disables the feature with a warning; the graph timing path is not a UI/profiling overlay by itself.

Production completion criteria

1. Define one non-blocking frame boundary that advances CPU profiling and memory sampling on the correct owning thread.
2. Publish a copied telemetry snapshot to ZUI/debug tooling; never let UI read a live mutable profiler buffer concurrently with its writers.
3. Surface dropped CPU samples, query-pool capacity exhaustion, unavailable queue timestamps, and arena registrations as explicit telemetry.
4. Verify CPU and GPU timing under all supported backends and device-lost, resize, and no-timestamp-feature cases. Timing must never add a GPU wait to normal presentation.
5. Establish build/release policy for profiling symbols, Tracy enablement, capture privacy, and retention. Do not claim “zero overhead” for code paths that still compile instrumentation under ZENGINE_PROFILING.