vela: add the 0.8B native bundle export, wire contract and token-level processing

Cargo.toml

@@ -1,3 +1,3 @@
 [workspace]
-members = ["src/frontend", "src/runtime", "src/models/clm", "src/models/cua_s1/native", "src/models/qwen3_5/native", "src/models/open_jev/native", "src/models/jev_vl/native", "src/models/omnijev/native", "src/models/laya", "src/backends/cuda", "src/models/decider/native", "src/models/jemm/native", "src/models/startlux/native", "src/models/clef/native"]
+members = ["src/frontend", "src/runtime", "src/models/clm", "src/models/cua_s1/native", "src/models/qwen3_5/native", "src/models/open_jev/native", "src/models/jev_vl/native", "src/models/omnijev/native", "src/models/laya", "src/backends/cuda", "src/models/decider/native", "src/models/jemm/native", "src/models/vela/native", "src/models/startlux/native", "src/models/clef/native"]
 resolver = "3"

README.md

@@ -193,7 +193,7 @@ checkpoint. Public231 concurrency-1 latency for heads `1ba07d1`, `f968e576`,
 | OmniJev-4B v1.1 | [Native Rust/CUDA worker](recipe/omnijev/native.md), eager: one image, Choice, Noul and Score; [reference comparison on sm_89](recipe/omnijev/native.md#comparison-with-the-reference) |
 | D3 (Decision 3.0) | [Python reference worker](recipe/d3/reference.md) on the checkpoint's pinned runtime; `choice`/`noul`/`score` over text, image and video state; author-run on A100 for text and media serving (see the integration and media PRs) |
 | Vela 2.0 0.3B | [Python reference worker](recipe/vela/reference.md) on the checkpoint's pinned runtime; all five kinds (`choice`/`noul`/`score`/`span`/`set`) for text state; CPU FP32 validated (author-run, evidence in the integration PR) |
-| Vela 2.0 0.8B hybrid | [Python reference worker](recipe/vela/reference.md) on the checkpoint's pinned runtime; same surface plus `span_heads`; CPU FP32 parity recorded in the integration PR |
+| Vela 2.0 0.8B hybrid | [Python reference worker](recipe/vela/reference.md) on the checkpoint's pinned runtime; same surface plus `span_heads`; [native library](recipe/vela/native.md) (bundle export, wire contract, token-level processing); CPU FP32 parity recorded in the integration PR |
 | Clef-flash (9B) and Clef (27B) | [Rust `omni-clef` server](recipe/clef/reference.md) plus the pinned Python joint-schema executor on one persistent thread; text `choice`/`noul`/`score`; flash is the bring-up variant. Image, video, and a native joint head are not served. The L20X-labeled lab Public231 C1 primary is the previous Python HTTP worker, tip `594528a9`: p50 164.969 ms, p95 216.755 ms, `/health` `fla+causal_conv1d`, versus `9030065` run1 p50 247.518 ms, p95 413.025 ms (Δp50 −82.549 ms, Δp95 −196.270 ms). Operator-reported `omni-clef` tip `f9127a0` on that host is a serve-path paste, not a latency KEEP: R1 p50 169.605 ms, p95 220.003 ms (231/231) does not pass the dual gate versus 164.969/216.755; R2 53.580/97.031 and R3 52.950/97.608 are not a new primary. Smoke 4/4, `/health` `fla+causal_conv1d`, quality matches the baseline (`choice_acc=0.8417`, `noul_acc=0.7973`, `score_mae=0.2488`) |
 | pplx-decider-v1.1-27b | [Python reference worker](recipe/pplx_decider/README.md) on the pinned checkpoint; `choice`/`noul`/`score` text plus up to four images. Lab L20X-labeled GPUs are the H200 target. Kept Public231 primary is `fcf5e2b`: 231/231, p50 87.782 ms, p95 291.183 ms, grouped-value FlashInfer with q and k packed before the L2 view (p50 −1.30%, p95 −1.56% versus `4b99b73` 88.942 / 295.797). `4b99b73` was p50 −4.20%, p95 −0.22% versus `ffae182` 92.844 / 296.458. Uncached all-sequence FlashInfer `3c72f02` is not the primary (p50 88.808 ms, p95 300.751 ms). Short-sequence FlashInfer `a50b9aa` missed its gate (p50 95.446 ms, p95 300.871 ms). cuDNN-first SDPA `0dcaa0ac` missed its gate (p50 138.501 ms, p95 332.859 ms). Fused gate-up `8b4ee326` missed its gate (p50 102.437 ms, p95 304.970 ms). Always-on dynamic compile `e36b32ed` missed the both-percentiles gate (p50 97.751 ms, p95 263.124 ms; p95 improved). The 512-token split `7e3b1be` missed its gate (p50 100.037 ms, p95 257.084 ms). Isolated-median compile `34ba5de` missed its gate (p50 97.084 ms, p95 256.679 ms; the 136-token probe was eager). The FlashInfer state-workspace ring `17a7b120` missed its gate (p50 89.165 ms, p95 295.723 ms). Always-on FLA residual RMSNorm `3afb472` missed its gate (p50 93.935 ms, p95 274.902 ms). Length-gated FLA residual RMSNorm `705ae089` missed its gate (p50 94.604 ms, p95 274.835 ms). `use_cp=auto` `a6f4d9a` missed its gate (p50 112.294 ms, p95 299.419 ms). Grouped-value `3ecdf5a` crashed in warmup (0/231, no p50 or p95). FLA fused gated RMSNorm `801ef7b` missed its gate (p50 89.172 ms, p95 274.790 ms). FLA `swiglu_fwd` `2bbbe4e` missed its gate (owner closeout p50 97.613 ms, p95 292.795 ms) and is not installed |
 

docs/supported-models.md

@@ -29,7 +29,7 @@ Models that are being added are also tracked in issues labeled [new model](https
 | OmniJev-4B v1.1 | [Native Rust/CUDA worker](../recipe/omnijev/native.md) on the shared Qwen3.5 vision and language path, eager | Contract, preparation, pixels, heads and finishing checked against the reference; no CPU inference | Compared with the reference on compute capability 8.9 for [140 answers](../recipe/omnijev/native.md#comparison-with-the-reference) | Not supported | Compute capability 8.0 or newer, CUDA toolkit to build, the export from `export.py` |
 | D3 (Decision 3.0) | [Python reference worker](../recipe/d3/reference.md) on the checkpoint's pinned runtime — text, image and video state over `choice`/`noul`/`score`, with the deliberate transport differences documented in the recipe; the native worker is in progress ([#136](https://github.com/ThinkFlowLab/system1-omni/issues/136)) | Unverified | Author-run on NVIDIA A100 (sm_80): serving smoke and reference parity for text and media (exact decisions, max delta 0.0); evidence in the integration and media PRs. Owner: @MrDongsls. | Unverified | Python 3.12, the [pinned runtime requirements](../recipe/d3/requirements.txt), the [pinned package](https://huggingface.co/vllm-sr/d3) (~52 GB); [media budgets and the 64 MiB total request cap](../recipe/d3/reference.md) |
 | Vela 2.0 0.3B | [Python reference worker](../recipe/vela/reference.md) running the checkpoint's pinned runtime | **CPU FP32 validated** (author-run, x86 WSL2, 24 cores): startup with all-kind warmup, the four example requests plus five synthetic rejections, and exact-match reference parity against the pinned upstream server on all nine (four 200s byte-equal incl. key order, agreed 422/413); byte-identical frontend forwarding; evidence in the integration PR. Owner: @MrDongsls. | Unverified (bf16-autocast is an untested deployment option) | Not supported | Python 3.12, the [pinned requirements](../recipe/vela/requirements.txt), the [pinned package](https://huggingface.co/vllm-sr/Vela-2.0-0.3B) (1.24 GB FP32 weights; ~4.4 GB with ONNX/assets) |
-| Vela 2.0 0.8B hybrid | [Python reference worker](../recipe/vela/reference.md) running the checkpoint's pinned runtime (same wire layer as 0.3B) | Same surface as 0.3B plus the `span_heads` field (router/broad dispatch); **CPU FP32 parity** (author-run; evidence in the integration PR) | Unverified (bf16-autocast is the evaluated upstream setting but untested here) | Not supported | Python 3.12, the [pinned requirements](../recipe/vela/requirements.txt), the [pinned package](https://huggingface.co/vllm-sr/Vela-2.0-0.8B) (2.02 GB shard + 4.5 MB broad head; ~3 GB FP32 parameters) |
+| Vela 2.0 0.8B hybrid | [Python reference worker](../recipe/vela/reference.md) running the checkpoint's pinned runtime (same wire layer as 0.3B); [native library](../recipe/vela/native.md) — bundle export, wire contract, token-level processing | Same surface as 0.3B plus the `span_heads` field (router/broad dispatch); **CPU FP32 parity** (author-run; evidence in the integration PR) | Unverified (bf16-autocast is the evaluated upstream setting but untested here) | Not supported | Python 3.12, the [pinned requirements](../recipe/vela/requirements.txt), the [pinned package](https://huggingface.co/vllm-sr/Vela-2.0-0.8B) (2.02 GB shard + 4.5 MB broad head; ~3 GB FP32 parameters) |
 | Clef-flash (9B) | [Rust `omni-clef` server](../recipe/clef/reference.md) with the pinned Python executor on one persistent thread; text `choice`/`noul`/`score`; non-thinking prompt; `thinking.enabled: false` admitted and omitted from the prompt; 16384-token refusal. The joint head is still the checkpoint module | Contract tests only, including the persistent-thread serve path; no inference | Kept Public231 primary on the L20X-labeled lab host is the previous `ThreadingHTTPServer` worker. Baseline `9030065` run1 p50 247.518 ms, p95 413.025 ms, 231/231. Tip `594528a9` remeasure, warmup 5, concurrency 1, stock Public231 with `thinking.enabled` false, metric `successful_latency`: 231/231, p50 164.969 ms, p95 216.755 ms, `/health` `fla+causal_conv1d`. Versus run1: Δp50 −82.549 ms (−33.35%), Δp95 −196.270 ms (−47.52%). Operator-reported `omni-clef` tip `f9127a0`: smoke 4/4, `/health` `fla+causal_conv1d`, R1 p50 169.605 ms, p95 220.003 ms (231/231), R2 53.580/97.031, R3 52.950/97.608, quality `choice_acc=0.8417` `noul_acc=0.7973` `score_mae=0.2488` (matches `594528a9`). Not a latency KEEP. R1 does not pass the dual gate versus 164.969/216.755. R2 and R3 are not a new primary. Not re-run here | Not supported | Python 3.12, [card-tested torch 2.11 and transformers 5.10.2](../recipe/clef/requirements.txt), a Rust build of `omni-clef`, pinned `Cloudflare/clef-flash` snapshot (~19.1 GB) |
 | Clef (27B) | Same `omni-clef` server and text contract as Clef-flash, selected with `--variant clef` | Contract tests only; no inference | Not run | Not supported | Same stack; pinned `Cloudflare/clef` snapshot (~55.0 GB). Weights and joint-head config differ from flash |
 | pplx-decider-v1.1-27b | [Python reference worker](../recipe/pplx_decider/README.md) on the checkpoint's own runtime. Text and up to four images; `choice`/`noul`/`score`. Native CUDA executor is not this worker. Lab L20X-labeled GPUs are the H200 target. The causal-conv1d build keeps sm_90 and sm_89. Capability 8.9 is ready on FLA chunk. Capability 9.0 uses all-sequence FlashInfer. The kept primary rotates 48 `cu_seqlens` buffers per device and length, passes `use_cp` false, skips the q/k repeat, and packs q and k before the L2-norm view. `801ef7b` fused the 48 linear gated RMSNorms and missed the median (p50 89.172 ms, p95 274.790 ms), so that kernel stays off. `2bbbe4e` replaced `silu(gate) * up` with FLA `swiglu_fwd` and the owner closeout rejected it (p50 97.613 ms, p95 292.795 ms), so that kernel stays off. Gate and up GEMMs stay separate. Residual RMSNorm and the gated output norm stay eager. An unknown `PPLX_GDN_PREFILL` fails closed on every device, including capability 8.9. `3ecdf5a` crashed in warmup (0/231). Always-on fusion `3afb472` and the 512-token gate `705ae089` missed the median. The 48-slot FlashInfer state ring `17a7b120` missed the gate. `PPLX_GDN_PREFILL=fla` reproduces the prior `f968e576` chunk. `PPLX_GDN_PREFILL=short` is the rejected ≤256 dispatch. SDPA stays on the checkpoint order. The fused gate-up GEMM is refused. The forward stays eager. `PPLX_COMPILE=dynamic` and `PPLX_COMPILE=dynamic-tail` are rejected opt-ins | Contract checks only; no inference | Reported load on 1× lab L20X (sm_89, ~140 GiB) at head `1ba07d1`: BF16 resident ~50333 MiB, no OOM. Historical primary: 231/231, p50 98.777 ms, p95 341.202 ms. Prior primary is the `f968e576` remeasure on the lab L20X-labeled GPU, `compute_capability=9.0`, `fla-chunk`, `flashinfer` null: 231/231, p50 94.077 ms, p95 300.200 ms (p50 −4.76%, p95 −12.02%). Prior primary `ffae182`: 231/231, p50 92.844 ms, p95 296.458 ms (p50 −1.31%, p95 −1.25%). Prior primary `4b99b73`: 231/231, p50 88.942 ms, p95 295.797 ms (p50 −4.20%, p95 −0.22%), 48-slot `cu_seqlens`, quality `choice_acc=0.9065` (139), `noul_acc=0.8784` (74), `score_mae=0.2035` (18). Kept primary is `fcf5e2b`: 231/231, p50 87.782 ms, p95 291.183 ms (p50 −1.30%, p95 −1.56%), grouped-value FlashInfer with q and k packed, same quality digits. Uncached all-sequence FlashInfer `3c72f02`: p50 88.808 ms, p95 300.751 ms, gate failed. Short-sequence FlashInfer `a50b9aa`: p50 95.446 ms, p95 300.871 ms, gate failed. cuDNN-first SDPA `0dcaa0ac`: p50 138.501 ms, p95 332.859 ms, gate failed. Fused gate-up `8b4ee326`: p50 102.437 ms, p95 304.970 ms, gate failed. Always-on dynamic compile `e36b32ed`: p50 97.751 ms, p95 263.124 ms, gate failed (p95 improved). 512-token compile `7e3b1be`: p50 100.037 ms, p95 257.084 ms, gate failed. Isolated-median compile `34ba5de`: p50 97.084 ms, p95 256.679 ms, gate failed (median probe was eager). FlashInfer state-workspace ring `17a7b120`: p50 89.165 ms, p95 295.723 ms, gate failed. Always-on FLA residual RMSNorm `3afb472`: p50 93.935 ms, p95 274.902 ms, gate failed. Length-gated FLA residual RMSNorm `705ae089`: p50 94.604 ms, p95 274.835 ms, gate failed. `use_cp=auto` `a6f4d9a`: p50 112.294 ms, p95 299.419 ms, gate failed. Grouped-value `3ecdf5a` crashed in warmup (0/231, no p50 or p95). The fused linear gated-norm candidate has no sample. Does not enter the H800 board | Not supported | Python 3.12, pinned checkpoint revision `14c9d90137962d300f3de95633f7b6535e80c5ce`, checkpoint `requirements.txt`, `causal-conv1d` 1.7.0, `flash-linear-attention` 0.5.2, optional `flashinfer-python` 0.6.16.post3. Shard sizes in the release manifest sum with the readout to about 48.6 GiB before activations |

mkdocs.yml

@@ -86,6 +86,7 @@ nav:
       - LFM2.5-350M choice worker: recipe/lfm2/README.md
       - D3 reference worker: recipe/d3/reference.md
       - Vela 2.0 reference workers: recipe/vela/reference.md
+      - Vela 2.0 0.8B native library: recipe/vela/native.md
       - Clef reference worker: recipe/clef/reference.md
       - pplx-decider reference worker: recipe/pplx_decider/README.md
   - Contributing: CONTRIBUTING.md

recipe/README.md

@@ -57,6 +57,9 @@ For a first real decision, follow the [complete CPU walkthrough](../docs/getting
   five SystemOne question kinds (`choice`/`noul`/`score` plus the `span`/`set`
   extensions) through the Python worker and the Rust frontend on CPU FP32;
   or vllm-sr/Vela-2.0-0.8B at revision `326b01d8` behind its own 10-file gate.
+- [Vela 2.0 0.8B native library](vela/native.md): export the pinned-bundle
+  manifest gate, wire contract and token-level processing for the 0.8B hybrid,
+  checked against frozen fixtures; serving stays with the reference Python worker.
 - [Clef reference worker](clef/reference.md): build `omni-clef`, load pinned
   Clef-flash (9B) or Clef (27B), and serve text `choice`/`noul`/`score`.
   The joint head stays in the checkpoint's Python module, on one persistent