PMAT-3477: Qwen3.5 4B/9B/27B on CPU and CUDA — block_count loader fix + Gated DeltaNet GQA (tiled head mapping)

[noahgift]

Part of EPIC #3477 (0.68.2 interrupt train). Operator ruling 2026-09-19: dense Qwen3.5 up to 27B on CPU+CUDA is required for 0.68.2.

Root causes

1. load_qwen35_layers read a bare block_count key that no GGUF carries → every Qwen3.5 file got 24 layers. 0.8B/2B have 24 so they were right by accident; 4B (32) and 9B/27B ran truncated models — that is QE2E-INV-001 asserts Qwen3.5-9B ∈ [9.0B, 9.2B] but the repo's own 9b descriptor sums to 8.209B — the descriptor declares no DeltaNet tensors #3346. Now reads {arch}.block_count and refuses when absent.
2. Gated DeltaNet GQA (4B/9B num_v_heads 32 vs num_k_heads 16, 27B 48/16): value head h reads key head h % num_k_heads — ggml's TILED order (llama.cpp's converter permutes V heads so ggml_repeat equals HF repeat_interleave). Applied in the CPU recurrence and the CUDA delta-rule kernel; recurrent state is [num_v_heads · Dk · Dv].

Measured (lambda, RTX 4090)

• cargo test -p aprender-gpu --lib --features cuda delta_rule 9/9; cargo test -p aprender-serve --lib --features cuda qwen35_cuda 16/16 (4B real-file per-layer parity worst 2.25e-2 < 5e-2; e2e cosine ≥ 0.9995, argmax equal at every position).
• 4B raw-prompt greedy tokens [11751,13,198,32] == llama.cpp (pinned test); apr run --gpu 4B 7.7 s / 9B 15.0 s / 27B 34.6 s → "Paris." on the CUDA backend line, no OOM; --no-gpu 4B/9B/27B → "Paris.".
• Falsification run: mapping h / ratio turns 4 grouped device shapes RED; kernel rem → div fails the 4B layer test (6.3e-1) and e2e argmax at pos 0.
• Two-host ladder at this head (scripts/model_ladder.sh, 8 rungs, cpu+cuda golden per rung): lambda (RTX 4090 sm_89) receipt lambda:/mnt/nvme-raid0/agent-wt/rel-0682-autopilot/receipts-c58062517/lambda.json — qwen2-1.5b, qwen3-1.7b, qwen35-0.8b/2b/4b/9b/27b OK "backends cpu,cuda honoured"; only the optional qwen3-8b rung red (golden "Empty output", 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, unchanged from 0.68.1). gx10 (GB10 sm_121) at the content-identical pre-rebase head 6396fdd: 8/8 green incl. GPU golden on all five Qwen3.5 sizes (gx10:~/probe/receipts-6396fdd66/gx10.json); re-run at c580625 (this head): gx10:/home/noah/probe/receipts-c58062517/gx10.json — executed=8 red=0, every rung "qa cap+golden pass, backends cpu,cuda honoured" (incl. qwen3-8b), the five Qwen3.5 rows carrying golden_output "3 golden test cases passed through the runtime entry point (GPU hybrid forward, GPU: implement Gated DeltaNet / SSM inference for Qwen3.5 / Qwen3.8 hybrid GGUFs #3090)". Receipts are NOT committed here (evidence/dogfood/models/0.68.1/ is owned by evidence(ladder): 0.68.1 receipts re-measured at main a6f54e84c on both hosts — eight rungs with sha256, RED by construction on Qwen3.5 (#3510) #3523; the 0.68.2 receipts land in the bump PR at T-2).
• Review quorum 3/3 PASS: docs/audits/quorum-PMAT-3477-pr3527.json.

keep-open: #3477 (epic closes at the tag)
keep-open: #3090 (closes when the ladder receipts land)
Closes #3346
Closes #3510

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3527 head=f36aa1bea21fd369c186801fbdba0bb8291ea104 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.