This ledger records which upstream reports informed the libfreenect2-metal 0.3 work. “Covered” means the fork contains the documented implementation and hosted regression coverage; it does not change the state of the original OpenKinect issue. Hardware-only checks remain release gates where noted.

Upstream issue	v0.3 coverage	Validation
#792	C++11 standard threading; monotonic arrival timestamps; bounded timestamp-aligned listener; replay timing	timed-wait, wraparound, queue, and interruption unit tests
#438	immutable calibration snapshot; durable manifest-directory raw recording; strict/salvage replay; recording calibration inputs	round-trip, truncation, traversal, replay-selection, timing, and parser-fuzz tests
#381	standard threading; idempotent failure cleanup; device state/error reporting; explicit wait/reopen recovery	lifecycle unit coverage; unplug/replug hardware soak required
#163	independent bilateral and edge-aware controls on CPU, OpenCL, CUDA, OpenGL, and Metal	four-mode CPU/Metal tests plus hosted OpenCL/OpenGL matrix
#721	first-transfer host arrival time and explicit timestamp-aligned software pairing	fragmented RGB, completed-depth, wrap-safe comparison, and listener tests
#869	documented separation between timestamp pairing and hardware synchronization	timing and capture metadata tests
#978	context-bound idempotent OpenGL teardown and partial-initialization cleanup	construction/teardown stress test; OpenGL/hardware soak required
#948	raw color/depth recording with identity, calibration, metadata, ordering, and replay	synthetic CPU round trip and malformed-input fuzzing
#1001	explicit VAAPI device selection and deterministic JPEG-capable node probing	Linux VAAPI warnings-as-errors build and fake initialization tests
#1071	VAAPI resource lifetime separated from decoder health	failure-path compile and unit coverage
#1123	packet buffers remapped on VAAPI failure paths	simulated runtime failure and same-packet fallback test
#1172	per-pipeline decoder policy, environment precedence, and permanent TurboJPEG fallback	API precedence and failover tests; Intel/AMD runtime soak required
#1191	unsafe NVIDIA nodes skipped only during automatic VAAPI probing	deterministic discovery build coverage; explicit-path behavior documented
#1173	project-owned CUDA vector primitives	CUDA 12.3 compile-only gate
#1190	standard and KDE kernels no longer include helper_math.h	both kernels compiled in the CUDA 12.3 container
#1202	CUDA sample-header discovery removed from modern and legacy CMake paths	GPU-less explicit-architecture build; hardware parity required
#144	per-device offset-only or linear depth correction with no automatic/global offset	synthetic known-distance and holdout tests
#319	explicit ROI statistics and guidance separating scalar Z correction from spatial calibration	invalid/outlier ROI tests and calibration guide
#535	live warm-up discard and thermal warm-up guidance	hardware-free collector build; warmed-device validation required
#596	serial/firmware-bound versioned profiles with mismatch refusal	JSON round trip and capture policy build coverage
#865	opt-in correction preserving invalid pixels and raw recordings	invalid-pixel, fit, replay-input, and holdout tests

Release-only hardware evidence

Before publishing v0.3.0, attach results for the available Kinect lifecycle soak, 60-second raw recording/replay validation, Intel or AMD VAAPI runtime fallback run, CPU/CUDA recording parity, and three-distance plus unseen holdout calibration. Hosted CI intentionally does not report GPU or USB runtime success when it only compiled or simulated those paths.

Table of Contents

Release-only hardware evidence