fix(perf): avoid replaying terminal B1 decode

[inureyes]

For a fused-lookahead-eligible capped B=1 request, the scheduler submitted the next lookahead at the length boundary, discarded both speculative appends, and synchronously replayed a token it had already sampled. This changes the terminal B=1 path to commit that pending token through the existing finish/finalization logic, removing two complete decode forwards per affected request while leaving B>1 behavior unchanged. Closes #2227.

Profiling on GB10 showed row plumbing (0.0053%), metrics (0.0462%), provider channel send (0.0698%), and lookahead collect (0.0836%) below the issue's 0.1% threshold, so those candidates remain unchanged. A guard mutation makes the live-lookahead test fail with four forwards instead of two, confirming that the new branch accounts for the removed work.

Performance, five interleaved quiet-host rounds against pre-epic 4c44e317:

Model	Prompt	Server dense vs baseline CLI	Baseline null arm
Qwen3-1.7B 4-bit	256	+0.57% (-0.25%..+1.54%)	+0.04% (-0.41%..+0.76%)
Qwen3-1.7B 4-bit	8192	+0.32% (+0.04%..+0.63%)	+0.22% (-0.13%..+0.66%)
Llama-3.2-1B 4-bit	256	+0.48% (+0.19%..+1.41%)	+0.07% (-12.38%..+0.54%)
Llama-3.2-1B 4-bit	8192	-0.16% (-1.15%..+0.05%)	-0.17% (-0.29%..+0.14%)

All four medians pass the -1.0% threshold. Every candidate range overlaps its null range, so the measurements establish the threshold without claiming a speedup. The unexplained Llama 256 null outlier (-12.38% in round 3) remains in the raw data and analysis without exclusion or rerun.

Validation:

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --features cuda -- -D warnings
• restored full CUDA suite: 8,930 root tests and 2,072 core tests passed; all integration binaries had zero failures
• repository contract checks
• four live-lookahead terminal scheduler tests, including max-token boundaries, stochastic sampling, EOS precedence, and unchanged B=2 behavior
• controlled guard mutation: expected assertion failure, actual 4 forwards versus expected 2
• real mlxcel run before/after generated text: byte-identical
• Llama full engine parity: 13 identical, zero divergent, one not applicable
• Qwen no-prompt-cache engine parity: nine identical, zero divergent, one not applicable

The full Qwen parity command still exits 1 on its prompt-cache miss/hit pair at greedy token 58 (374 versus 594) and seeded token 61 (576 versus 362). Forced synchronous decode reproduces the same positions and token IDs, while the new guard is unreachable; the no-cache pairs affected by this change are identical. The benchmark report records this existing prompt-cache limitation, which is outside this issue's scope.

Evidence: benchmark report and raw data.