[noahgift]

Closes #3090.

Operator decision 2026-09-19 (in-session, verbatim selection): "GPU required — build #3090 now". This PR is the GPU half of the 0.68.2 interrupt (EPIC #3477); the CPU half landed in #3484.

What lands

commit	what	falsifier
5ebc79d	six Gated-DeltaNet PTX kernels in crates/aprender-gpu/src/kernels/gdn/ (causal conv1d+SiLU, per-head L2 norm, dt/β gates, delta-rule recurrence, gated RMSNorm, sigmoid gate)	device tests vs a verbatim host port of the CPU fns at real 0.8B shapes; delta-rule mutant (scale×1.01) RED at relative L∞
7591ff1	head_dim-256 decode attention (positions-per-pass cap, multi-pass tested), partial NEOX RoPE (θ-scale on host — device pow measured 2.3e-4 > 1.4e-4 bar), [q|gate] split	12 device tests; stubbed kernels RED
58870fd, 6474c45, 1ed0adf	executor wrappers (graph-recording push on every launch) + Qwen35CudaModel: DeltaNet and attention layers op-for-op, forward_single, config-sized device state, download_layer	every DeltaNet/attention layer teacher-forced vs CPU; e2e 6-token: argmax identical at every position, cosine ≥ 0.996 (measured 0.9983–0.9996), per-op ≤ 1e-3 vs an exact-arithmetic reference; sigmoid-gate mutant RED (cos 0.766); DP4A falsifier RED (cos 0.466…−0.25)
d9459c7	refusal sites lifted in lockstep: capability gate admits qwen35 on GPU under cuda, parity/ptx-map stop refusing it, quant whitelist judges the SSM tensors, SSM predicate names the hybrid path	28 unit tests across apr-cli/serve
b2bd0c5	apr run/chat/qa route qwen35 to the GPU; F2 guard judges the hybrid against its own CPU forward (Serial probe); no silent CPU fallback; stale "GPU declined (#3090)" skip reasons removed	apr run --gpu: CUDA backend line, "Paris"; apr qa: ALL GATES PASSED, Golden Output + Throughput on the GPU

Two measured facts that shaped it

• The CPU production forward quantizes activations to Q8_K (2.95e-3 per projection from exact arithmetic; the GPU float GEMV is 1e-7), so "device == CPU at 1e-3" is not satisfiable by a correct GPU — the contract is restated in contracts(qwen35-gpu): QHF-GPU-008 judged against exact arithmetic, QHF-GPU-009 end-to-end argmax + cosine floor vs production (#3090) #3514 (per-op exact ≤ 1e-3; e2e argmax + cosine ≥ 0.996). A floor test (qwen35_cuda_the_parity_floor_is_the_cpu_references_own_activation_quantization) bites in both directions.
• The production DP4A Q4K/Q6K GEMV profile is catastrophic through the recurrence (garbage argmax) while the float MWV variant is exact; Qwen35CudaModel pins the float variant (DP4A Q4K/Q6K GEMV profile is catastrophic through the Gated DeltaNet recurrence — Qwen3.5 on sm_89 needs the float MWV variant (PMAT-806 class) #3513 for the root cause).

Measured on lambda-vector (RTX 4090), pinned binary

Backend: GPU (CUDA, NVIDIA GeForce RTX 4090, 24035 MB VRAM) [qwen35 hybrid forward, #3090]
Output: The capital of France is **Paris**.
apr qa: Capability Match PASS · Golden Output PASS (GPU) · Throughput 50.5 tok/s PASS (GPU) · PTX Parity 6/6 · ALL GATES PASSED

gx10 (GB10) + ladder receipts: peer session runs both hosts at b2bd0c5; receipts land with the bump PR (#3498).

Process

Design grill 3/3 (refute-as-written → all four blocker classes applied: head_dim 256, partial rope, F2 reference, lockstep refusal sites). Pre-merge review quorum: in flight (docs/audits/quorum-PMAT-3477-pr.json to follow). Workers: G, H, I, J, J2, J3, K1, K2, K3 — every acceptance re-run by the orchestrator.

ont-delta: none this PR adds device kernels and a GPU forward; no ontology type, shape, reason or resolution changed

🤖 Generated with Claude Code

Issue refs (§6 R-2)

keep-open: #3477 — the release epic closes at the tag, not at this PR.
keep-open: #3513 — DP4A-through-recurrence root cause (0.69.0), mitigated here by pinning the float GEMV.
keep-open: #3346 #3510 — dense sizes 4B/9B/27B: the GQA-in-GDN CPU fix rides this branch; the CUDA half and the size rungs follow.

Two-host ladder at b2bd0c5 (peer session, main contract overlaid)

• gx10 (GB10, sm_121): rc=0, 4/4 rungs green; qwen35-0.8b-q4km cuda {ran:true, fallback:false}, golden PASS — ~/probe/receipts-b2bd0c5dc-mainladder/gx10.json
• lambda (RTX 4090, sm_89): qwen35-0.8b-q4km cuda ran, no fallback, golden PASS; only the optional qwen3-8b rung red (apr qa Golden Output reports 'Empty output' on Qwen3-8B — 32-token budget is consumed by <think> and the stripper leaves nothing; the rung cannot be certified by the harness #3486) — /home/noah/probe/receipts-b2bd0c5dc-mainladder/lambda.json

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3517 head=0f5aaab225105afa3aff64d977d18e91c7d84473 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.