# Sampler cache admission and retry NVN samplers pass through an ID/generation table, a thread-local parameter/handle cache and a global owning cache. Ignoring `vkCreateSampler` failure could admit an undefined error-output handle, or retain null in those caches indefinitely. Checking only the factory result would still let the outer ID table retain null. Only successful handles now enter the owning or acceleration caches. The ID table does not mark a failed creation resolved for that registration generation. Failure of one key is bounded to one attempt per present; later presents can retry the same ID, including on a fresh thread. Successful hits and unrelated keys remain usable. Failure keys are cleared at a later present. If failure tracking itself cannot allocate, new keys wait until the next present, while already-owned successful handles still work. A newly created sampler whose global cache insertion throws is destroyed before returning null. An error-output handle is never returned or destroyed. The parameter and thread-local handle tables are acceleration caches: allocation failure there does not discard registered parameters or destroy a valid global handle. Existing registration/version checks, immutable successful sampler lifetimes, filter/LOD/compare/anisotropy/border/reduction policies and the renderer API are unchanged. Existing replay locking serializes the shared cache and failure tracking; no new worker or GPU wait is added. Before either pushing descriptors or allocating/updating a descriptor set, `bind_descriptors` rejects a combined-image-sampler array containing a failed sampler. It leaves the pushed-state cache untouched and returns failure to the existing draw, compute and mesh-fallback callers. A later rebuild can retry; no replacement sampler changes the requested filtering policy. Storage images continue to accept their unused null sampler. `tests/test_sampler_admission.py` executes the complete actual sampler factory, ID-cache consumer and parameter policy helpers, using the public NVN parameter layout. Vulkan, registration/environment/hash inputs and the replay-lock boundary are modeled. Standard containers use tagged throwing allocators to target each cache operation. The owning-cache fault is armed after successful resource creation, independent of an STL's sentinel/bucket allocation choices. The 108 cases cover failure/retry/cache chains, ownership and allocation rollback, independent keys, registration changes, threads and unchanged sampler policies. The actual complete descriptor-binding function and fingerprint helpers run against modeled Vulkan calls too: cold and warm failures in an array's second element, graphics/compute, pooled sets, both pushed-state layouts, full pushes, next-present recovery and storage images. Rejected writes must produce no allocation, update, bind or push and must preserve the prior fingerprint. Run `python -I -B tests/run_all.py sampler_admission --strict`, or invoke the test with `--headers --cxx `; `--sanitize` is supported on Linux. Game-free failure models do not prove native OOM/device-loss recovery, image-quality changes, a full campaign or performance improvements.