feat(deepseek4): add DSpark speculative runtime

[davide221]

Stack

Depends on #503, which depends on #502:

#502 ROCmFPX ggml -> #503 DeepSeek4 target -> #496 DSpark runtime

This branch was rebuilt directly on #503 so it does not carry or conflict with older copies of the target and fused-decode changes.

Summary

• bridge DeepSeek4/MTP DSpark auxiliary heads into the draft GGUF contract
• add the DeepSeek4 DSpark loader, proposal graph, target feature capture, batched verification, and rollback loop
• add confidence-based adaptive verification width while retaining fixed-width controls
• keep standard AR unchanged unless DFLASH_DS4_SPEC is enabled

Validation

• Release HIP build for gfx1151;gfx1201
• R9700 DSpark load + forward at context 42: 20,480 finite floats, RMS 0.2820
• fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 optimizations #503 fused AR path remains selected when DFLASH_DS4_FUSED_DECODE=1
• coherent greedy generation with local and remote DSpark execution
• git diff --check

Follow-up

Remote DSpark execution and shared transport are isolated in #505.

[davide221]

Validation update (2026-07-09): the converter/loader bridge in this PR was validated on lucebox2 with the converted DSpark drafter artifact. In the follow-on DeepSeek4 runtime worktree, DSpark loads successfully, prefill feature capture and generation are wired, the batched-verify causality bug is fixed, and causal batched output matched exact sequential verification in paired probes. Adaptive EWMA width and fused batched verify are working; fused verify reached approximately 20 tok/s (peak 20.1 tok/s), versus roughly 13–15 tok/s for the earlier dynamic causal path. Exact sequential verification measured 1.78–1.91 accepted candidates per 4-candidate step. These runtime changes are follow-on work in the hub-494 experiment worktree and are not claimed as already included in PR 496.

[davide221]

Updated DSpark adaptive width in 1fc4f00.

• Fused Markov generation now returns confidence scores in the existing token-id synchronization (no extra host round trip).
• A compatible confidence head automatically selects q=2/q=3/q=4 from cumulative prefix confidence.
• Artifacts without a compatible confidence head fall back to the existing EWMA controller.
• No new environment variables or rejected confidence+EWMA hybrid policy were added.
• Expert k=3 is not included.

Validation on lucebox2 / gfx1151 with the exact committed source:

• HIP dflash_server build: pass.
• GSM 5/5 + MATH 5/5: 10/10.
• Server-side decode: 1,485 tokens / 50.769 s = 29.25 tok/s weighted.
• Fixed-q=4 reference: 29.49 tok/s weighted.

The confidence policy therefore retains essentially all high-acceptance throughput while adapting verification width on harder prompts.