fix(wgpu-parity): multi-step gate catches 7B Q4K autoregressive drift (#1864 wgpu)

[noahgift]

Summary

Closes the wgpu side of #1864 (the cuBLAS FP8 side stays open pending bisect).

apr run on Qwen2.5-7B Q4K via wgpu produced ampiezza gibberish with exit 0. Two loopholes — both now closed:

Loophole	Where	Fix
GGUF wgpu path had NO parity gate at all	try_wgpu_generate:61	Add multi-step gate (matches APR path)
Single-step gate missed autoregressive drift	try_apr_wgpu_inference:455	Extend to multi-step (3 by default)

Five-whys

1. apr run on 7B Q4K wgpu produces ampiezza with exit 0.
2. Init-time parity gate (FALSIFY-CPU-GPU-005) compares only step 0; cosine ≥ 0.99 there.
3. Subsequent autoregressive steps drift because the wgpu KV cache accumulates error.
4. The single-step probe couldn't observe drift because it never advanced past step 0.
5. Root cause: the gate's domain was too narrow (one step instead of N).

Fix (FALSIFY-CPU-GPU-006 / multi_step_parity_gate)

For N=3 steps (default; override via APR_WGPU_PARITY_STEPS in [1, 16]):

• Run CPU forward_single_with_cache + wgpu fwd.forward_layer, advancing both through the same deterministic token sequence (CPU argmax on cpu_logits steers both).
• Cosine-compare full vocab logits at every step.
• ANY cosine < 0.99 emits WGPU_FALLBACK_LOG_PREFIX tagged at step <k>/<N> + returns Err → caller falls back to CPU.

Single-step (N=1) is backward-compatible by construction (the v1.3.0 gate).

Live discharge (RTX 4090, this host)

$ apr run /home/noah/models/qwen2.5-coder-7b-instruct-q4_k_m.gguf \
    "What is 2+2?" --max-tokens 16

Backend: wgpu (Vulkan)
[PMAT-333] Dequantized 337 weights, 28282.5 MB F32
[apr-cpu-vs-gpu-output-parity-v1] wgpu path rejected,
  attempting fallback: cosine vs CPU = 0.722249 (< 0.99) at step 1/3
Output:
2 + 2 equals 4.    ← correct CPU output (was: "ampiezza")

Pre-fix on the same host/model: ampiezza\nampiezza (per #1864).
Post-fix: correct CPU answer + clear tagged rejection log.

Contract

apr-cpu-vs-gpu-output-parity-v1.yaml → v1.6.0:

• new equation multi_step_parity_gate
• new falsifier FALSIFY-CPU-GPU-006 (LIVE_DISCHARGED 2026-05-22)
• algorithm_evidence includes the full smoke trace above

Scope

This closes the wgpu visibility + autoregressive-drift loophole — no more silent gibberish reaches users. The underlying wgpu kernel bug that produces drift on hidden_dim=3584 (the actual numerical correctness issue in attention or FFN dispatch) remains open as a #1864 wgpu-kernel sub-issue. Until that root-cause lands, 7B Q4K wgpu correctly falls back to CPU at init.

The cuBLAS FP8 path (<|im_start|> gibberish on apr qa Golden Output) is a SEPARATE regression and stays open in #1864 pending bisection of v0.34.0..HEAD on cuda_chat_backend.rs.

Test plan

• cargo check -p aprender-serve --features cuda
• cargo test fallback_log_prefix — 3/3 PASS (backward compat)
• Live: 7B Q4K wgpu → CPU fallback → "2 + 2 equals 4." (was gibberish)
• FALSIFY-CPU-GPU-{001..005} backward compat (single-step reduces to v1.5.0 design)
• CI: workspace-test, contracts-lib, fmt

🤖 Generated with Claude Code