I ran a separate Windows Vulkan adapter on an NVIDIA GeForce RTX 4090 against the frozen compressed-merge source f4c26bd0b86e006e39581357afb0ee092f8da3f6. All 75 canonical cases passed twice in fresh processes, for 150 successful runs. Each output matched the complete canonical file bytes and hash; the input bytes and every allocation's prefix and suffix guards stayed intact. Vulkan API and synchronization validation were enabled, with zero warnings or errors. Source, shader, binary and fixture fingerprints stayed unchanged.

The qualification record retains each case's counts, output fingerprint and result, together with the original report and adapter identities. The baseline uses MSVC 19.44.35228.0 and portable CRC32C. The raw Vulkan driver version is 2559967232.

A separate host AddressSanitizer runner passed three selected cases: empty-key-tombstone, proper-prefix-and-binary, and boundary-mixed-257. Complete output, input preservation, guards and Vulkan validation all passed; the 43 artifact pins and 949 frozen-file pins stayed unchanged. The qualification record retains this runner's separate identity and results. This covers three host cases, not the entire corpus or shader-memory instrumentation. The exported evidence includes the report and pinned build configuration; I have not independently inspected the runner's binary instrumentation or raw sanitizer logs.

These are guarded correctness runs, not performance measurements. The adapter stages mapped input into device memory and reads the complete output back; it does not assume zero-copy host memory. Its eventual complete-path measurements must include allocations, initialization, transfers, scalar waits, output mapping and CRC. Pipeline compilation, fixture validation, fsync and catalog publication remain outside that construction measurement.

This qualifies the earlier frozen compressed implementation. It does not establish the performance of the current Metal optimizations on a discrete GPU, or qualify an integrated durable Vulkan backend. The Windows adapter is not yet part of this optional package. I keep its results separate from the Metal calibration.