test(cuda/vulkan): three *ParityTests classes cannot detect the bug they exist for (no type assertions; corpus-length-only; gate-region-only)

[jamesburton]

Three classes named *ParityTests that cannot detect the bug they exist for

C1 — zero quantisation-type assertions across ~20 named-quant end-to-end tests

tests/DotLLM.Tests.Integration/Vulkan/RealGgufVulkanParityTests.cs (tests at :75,91,111,138,157,169,187,215,234,246)
tests/DotLLM.Tests.Integration/Cuda/RealGgufCudaParityTests.cs (tests at :66,81,102,120,135,146,164,175,191,202)

grep -c QuantizationType returns 0 in both files. Each test is named <Model>_<QUANT>_VulkanForward_MatchesCpuReference and asserts only Architecture — trusting the filename and an env var. The doc at :162 explicitly invites substituting any GGUF via env var.

So a completely broken IQ1_S kernel passes when the env var points at a GGUF containing no IQ1_S tensor. The sibling file already warns about exactly this in prose: RealGgufQ3KDequantParityTests.cs:44-46 names a SmolLM-135M i1-Q3_K_M build that carries no Q3_K tensor at all. This is #344 mandate (4), violated in the files that most need it.

The tolerances are barely wider than a total scramble. :59-63 sets LogitsAbsTol = 3.0, TopKJaccardFloor = 0.5, StrictArgmaxFloor = 5/9. The same file records at :200-208 that #311 — a layout transposition producing weights with correlation 0.006 — measured L-inf 5.09 and Jaccard 0.40. Total garbage cleared the bound by 1.7x and 0.10. Any layout bug affecting a subset of sub-blocks fits inside. Constants duplicated at RealGgufCudaParityTests.cs:50-52.

C2 — a *ParityTests class whose only assertion is corpus length

tests/DotLLM.Tests.Integration/Cuda/CudaAttentionSplitKvGenerationParityTests.cs:123 and :314. The complete assertion set in both methods is Assert.True(fullCorpus.Length >= 1100). Every token, top1/top2 margin and both pre-/post-gate perplexities go to ITestOutputHelper and are never compared — not even a finiteness check.

Passes with attention_f32_split_kv returning zeros, un-softmaxed scores, a wrong hkv = hq % numKvHeads, or NaN, at every decode step at or above the gate.

C3 — GQA-split parity asserted only where the kernel is switched OFF

tests/DotLLM.Tests.Integration/Cuda/CudaAttentionMmaDecodeGqaSplitGenerationParityTests.cs:114,148,260. Three assertions total: corpus length, dispatched > 0, and Assert.Equal(0.0, maxPreGateStepDiff) — the last covering only depth < AttentionGqaSplitMinSeqKv, i.e. the region where the kernel under test is not called. Every post-gate step is log-only.

Passes for any implementation correct below the gate and wrong above it — specifically the kvHead = qHead % numKvHeads vs qHead / group swap, which is the kernel's entire purpose.

Fixes, all with in-repo precedent

• C1: lift the 5-line type-assertion block from RealGgufQ3KCudaForwardParityTests.cs:74-79 into RunGgufParityTest, taking the expected QuantizationType alongside expectedArch, with a tensor-count floor so one stray tensor cannot satisfy it. CrossBackendQuantGateTests.cs:204 is the stronger model — it requires every transformer-block weight to be the expected type.
• C1 tolerances: re-derive from a clean distribution and document the reasoning, as RealGgufQ5_0ParityTests.cs:44-64 does.
• C2: assert pplPostOn / pplPostOff <= 1.01 plus token-for-token agreement — the shape VulkanSplitDecodeMoeParityTests.cs:118,137 already uses.
• C3: assert post-gate steps, not just pre-gate.

Per #417, each fix must be demonstrated to fail against the wrong implementation named above.

Source: .docs/test-discrimination-audit.md sections C1-C3. Related: #417, #344, #311.