Add Nemotron Lightning target/MTP graph exports and explicit BF16 reconstruction

docs/gguf-capability-catalog.md

@@ -55,7 +55,7 @@ Runtime support above is independent from tokenizer materialization support belo
 
 | Evidence ID | Pinned candidate | Bounded result | Withheld runtime claims |
 |---|---|---|---|
-| `nemotron-h-moe-30b-iq2-xxs-runtime-blocker` | `bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF@1fc64d5b160654ec892df2708aa893b0e96e6491`<br>`nvidia_Nemotron-3-Nano-30B-A3B-IQ2_XXS.gguf`<br>18,010,755,296 B<br>`f3da710c046ce7cc6ff28a9b5f1a9153ac72e3f60603e51c7bb679d80716b58a` | config/tokenizer `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16@bf77c3174f68ad409e1c2aa60daeb46e32d1c606`; GGUF tokenizer metadata `6089bcaf08b3fe0d49379ca7e85bd3c93e8705bac6130636425c159212971225`; result=blocked; 401 tensors / 31,577,940,288 parameters; ORT 1.29.0 / CPUExecutionProvider; ort-genai 0.15.2; The smallest model GGUF in the pinned 29-file repository revision is 18,010,755,296 bytes, above the 16 GiB bounded-artifact policy; explicit float16/float32 weights require 63,155,880,576/126,311,761,152 bytes.<br>The normal optimized CPU export has 37,142 nodes (40,167 before Mobius optimization) and 6,028 MatMul nodes because its truthful ReLU2 routed experts remain an ONNX loop. ORT 1.29 MoE/QMoE exposes ReLU but not ReLU2. QMoE has separate router_probs/router_weights for correction-biased selection with unbiased sigmoid mixing, while shared experts and optional latent projections can surround it; those are not fused-op blockers. The pinned 30B has no latent projection. QMoE does not support GGUF IQ2_XXS storage, and Mobius has no proven IQ2_XXS packer/kernel path.<br>ORT GenAI 0.15.2 discovers sparse/nonconsecutive KV and conv/recurrent slots, but derives recurrent_state names while this export uses ssm_state, does not beam-reorder recurrent state, and rejects nonzero recurrent-state rewind.<br>The GGUF tokenizer declares pre=pixtral, whose compiled llama.cpp behavior is not serialized; exact ORT tokenizer materialization is unavailable. | cached decode and deterministic generation, full-logit parity, package and report roundtrip, state replay, rollback, and reorder, tensor value closure |
+| `nemotron-h-moe-30b-iq2-xxs-runtime-blocker` | `bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF@1fc64d5b160654ec892df2708aa893b0e96e6491`<br>`nvidia_Nemotron-3-Nano-30B-A3B-IQ2_XXS.gguf`<br>18,010,755,296 B<br>`f3da710c046ce7cc6ff28a9b5f1a9153ac72e3f60603e51c7bb679d80716b58a` | config/tokenizer `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16@bf77c3174f68ad409e1c2aa60daeb46e32d1c606`; GGUF tokenizer metadata `6089bcaf08b3fe0d49379ca7e85bd3c93e8705bac6130636425c159212971225`; result=blocked; 401 tensors / 31,577,940,288 parameters; ORT 1.29.0 / CPUExecutionProvider; ort-genai 0.15.2; The smallest model GGUF in the pinned 29-file repository revision is 18,010,755,296 bytes, above the 16 GiB bounded-artifact policy; explicit float16/float32 weights require 63,155,880,576/126,311,761,152 bytes.<br>The normal optimized CPU export has 40,222 nodes (43,272 before Mobius optimization) and 6,028 MatMul nodes because its truthful ReLU2 routed experts remain an ONNX loop with FP32 routing/accumulation. This graph-only census was refreshed locally with ORT 1.30.0 installed; the pinned ORT 1.29 runtime inspection below is historical, not a new runtime result. ORT 1.29 MoE/QMoE exposes ReLU but not ReLU2. QMoE has separate router_probs/router_weights for correction-biased selection with unbiased sigmoid mixing, while shared experts and optional latent projections can surround it; those are not fused-op blockers. The pinned 30B has no latent projection. QMoE does not support GGUF IQ2_XXS storage, and Mobius has no proven IQ2_XXS packer/kernel path.<br>ORT GenAI 0.15.2 discovers sparse/nonconsecutive KV and conv/recurrent slots, but derives recurrent_state names while this export uses ssm_state, does not beam-reorder recurrent state, and rejects nonzero recurrent-state rewind.<br>The GGUF tokenizer declares pre=pixtral, whose compiled llama.cpp behavior is not serialized; exact ORT tokenizer materialization is unavailable. | cached decode and deterministic generation, full-logit parity, package and report roundtrip, state replay, rollback, and reorder, tensor value closure |
 | `glm-dsa-iq1-s-artifact-budget-blocker` | `unsloth/GLM-5.2-GGUF@abc55e72527792c6e77069c99b4cb7de16fa9f23`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00001-of-00006.gguf` 9,423,744 B `46b6148389219ae45167cb8124fbb18ef7d432daf619b4faf9e06ea80d3f4777`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00002-of-00006.gguf` 49,208,128,256 B `f2180207285e04fcaa5b8c53ba6e77ad5cc58666b6e7c6b04a5eded3fe8bef09`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00003-of-00006.gguf` 49,684,417,024 B `b1c0c5a302cc8d5d9ea0bcd4467c01db72c26839f820f7e882079582ea0a8d2b`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00004-of-00006.gguf` 49,396,052,864 B `a6a42da6975e29f89866dcde2956e9e50e6ea26635fb5063b74f3973f4f863b6`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00005-of-00006.gguf` 49,246,275,936 B `a4a9851a50db533f21ef824e5d8038f04e6782e7d602d18e5fdd6643f68ccccb`<br>`UD-IQ1_S/GLM-5.2-UD-IQ1_S-00006-of-00006.gguf` 19,171,063,136 B `3b767f55df64e0432d52fcf1a14eb47a1ef3bbc91339e2ae220f38602237d7d7`<br>total 216,715,360,960 B | blocked; The smallest mainline GLM-5.2 GGUF is a 216,715,360,960-byte split set whose trunk also declares an unsupported routed DSA/MLA MTP block. | real-weight full-logit parity, runtime packaging, deterministic generation |
 | `minimax-m2-iq1-s-artifact-budget-blocker` | `mradermacher/MiniMax-M2-i1-GGUF@2d4f9b1a86d32ce4dfc47db312c8d6fcae8d7b37`<br>`MiniMax-M2.i1-IQ1_S.gguf` 46,514,882,176 B `7bae986e3cd380c28c6177d612fce1d52373241a0dfa13a6fca25de79abf15fb`<br>total 46,514,882,176 B | blocked; The smallest immutable public MiniMax-M2 GGUF is 46,514,882,176 bytes; real-weight parity cannot enter the 16 GiB bounded evidence set. | real-weight full-logit parity, runtime packaging, deterministic generation |
 | `mistral4-iq1-m-artifact-budget-blocker` | `unsloth/Mistral-Small-4-119B-2603-GGUF@bd93c721735aa32c035c0f19e738cb3371fd56ff`<br>`Mistral-Small-4-119B-2603-UD-IQ1_M.gguf` 32,306,941,632 B `40fcdee4869110938638c6b8bac253f442b196518d4623f4afdf4b885cd961c7`<br>total 32,306,941,632 B | blocked; The smallest immutable public Mistral4 GGUF is 32,306,941,632 bytes; real-weight parity cannot enter the 16 GiB bounded evidence set. | real-weight full-logit parity, runtime packaging, deterministic generation |

docs/index.md

@@ -53,4 +53,5 @@ ai-model-support-strategy
 :caption: Research
 
 research/testing-strategy-analysis
+nemotron-lightning-export
 ```