Fix DSpark greedy-identity break: replay accepted tokens via single-token decode

[nexus-cw]

Summary

Fixes #658: DSpark speculative decode did not preserve byte-identical greedy
(--temp 0) output vs. non-speculative decode on longer generations.

Root cause (see linked issue for full detail)

ds4_session_eval_dspark_speculative_argmax's accept-commit logic had a fast
path for full accepts (commit_drafts == draft_n) that committed the verify
batch's own batched-GEMM-derived compressor-frontier KV state
(ds4_gpu_matmul_f16_pair_tensor) directly. That state is numerically
distinct (standard FP16/BF16 GEMM non-associativity) from what ordinary
single-token decode's fused kernel
(ds4_gpu_matmul_f16_pair_compressor_store_tensor) would have produced for
the same token, so it silently drifted from a pure-decode trajectory until it
eventually flipped an argmax tie on longer generations.

The partial-accept path already had the correct mechanism: roll the
KV/compressor-frontier state back to the pre-verify snapshot, then re-decode
each accepted token one at a time through the same single-token kernel plain
decode uses (metal_graph_eval_token_raw_swa). It just wasn't used for full
accepts.

Fix

Minimal diff: removes the full-accept fast path entirely. Every accept (full
or partial) now falls through to the existing rollback+replay path
unconditionally, so post-accept KV state and logits are always re-derived
through the same code path plain greedy decode uses -- bit-identical by
construction, not by numerical luck.

No new code, no new flags, no dialect/quantization changes. Zero effect when
no drafter is configured (draft_n is never reached in that case).

 ds4.c | 22 +++++++++++++---------  (1 file changed, 22 insertions(+), 51 deletions(-))

Identity-test evidence

Resident IQ2XXS quant of DeepSeek-V4-Flash + a DSpark support GGUF (--mtp <support.gguf> --dspark), --cuda --temp 0, drafter-enabled vs no-drafter,
byte-for-byte diff, built from this branch (based on current main):

prompt	tokens	result
"Explain the TCP three-way handshake."	400	byte-identical
"Write a detailed explanation of how the TCP three-way handshake..." (longer prose prompt)	800	byte-identical

Both prompts were confirmed to diverge on unpatched main before this fix
(the second one is the original reproducer from the linked issue, extended
past its divergence point here). No divergence observed post-fix in either
case.

Overhead (stated plainly)

This fix makes every accepted token -- including full accepts, which
previously skipped the replay -- pay for one extra fused single-token
forward pass. On our hardware/pairing (NVIDIA GB10, resident IQ2XXS +
this drafter), that increases drafter-enabled generation slowdown vs.
no-drafter from about -5% to roughly -14% to -17% across a few prompts
(single runs, not a full statistical sweep). This is the honest cost of
correctness here: the alternative (a narrower fix that re-derives only the
compressor-frontier projection instead of the full forward pass) is possible
in principle but a larger, riskier change, and wasn't attempted in this PR --
happy to discuss as a follow-up if there's interest.

Testing

make cuda-spark: clean rebuild, zero warnings, on this branch (based on
current main).

🤖 Generated with Claude Code