[antono3]

Newly named graphical executables repeatedly spent hundreds of milliseconds compiling the same eight Vulkan pipelines. Add a per-user application pipeline cache: in the controlled GTX 1060 comparison, median platform startup fell from 981 ms to 195 ms and pipeline creation from 816 ms to 28 ms. The measurements include cache-file loading and use fresh executable names; these are local observations, not universal guarantees.

Missing/corrupt/stale caches and directory, driver or write failures preserve rendering. Validate the application checksum plus Vulkan vendor/device/UUID header before import; save through a process-specific temporary file and atomic replacement. An empty TT_PIPELINE_CACHE_FILE disables caching. Keep the existing C renderer ownership and allocator/Vulkan pins.

Add opt-in GPU timestamp intervals and scene/post-processing capture labels with frame-fence synchronization, availability checks, counter-wrap handling and cleanup. Add reproducible compute/renderer/cache benchmarks and raw results. OpenCL binary caching is intentionally not added: its measured warm gain was about 1.2 ms, while checked compute remains slower than CPU-only work for these particle batches.

Validation: GCC build and runtime tests pass with compiler fallback disabled; headless checksums match. GPU profiling and cache roundtrip, truncation, payload corruption, stale vendor/UUID, missing-parent writes, default-location reuse and blocked-directory fallback pass on hardware Vulkan under validation. CI exercises the profiler on Linux, cache persistence on Linux and Windows, and the existing package consumers. Documentation links/source anchors and distribution tests (5) pass.