fix(specprefill): avoid dense tail expansion within cache window

[Thump604]

Summary

• avoid forcing full-tail inclusion for RotatingKVCache when the prompt still fits inside the cache window
• add a regression test that pins the no-eviction vs eviction boundary

Why

On the Nemotron 120B live runtime path we cap unbounded attention caches with max_kv_size=65536, which means the SpecPrefill target cache is a RotatingKVCache even for prompts that still fit entirely inside the window. The previous sparse-prefill logic treated any RotatingKVCache as if eviction were already active and unioned in the full max_size tail. At 64K this effectively collapsed sparse prefill back into dense target work.

Validation

• pytest tests/test_specprefill_rotating_cache.py -q
• black --check --fast vllm_mlx/specprefill.py tests/test_specprefill_rotating_cache.py
• python -m py_compile vllm_mlx/specprefill.py tests/test_specprefill_rotating_cache.py

Live runtime evidence from the promoted stack:

• Nemotron 64K clean cold rerun before fix: dense 271.98s, sparse 395.37s
• Nemotron 64K clean cold rerun after fix: dense 277.89s, sparse 143.90s
• Nemotron 16K clean cold rerun after fix: dense 68.48s, sparse 36.26s
• recall stayed correct in all fixed runs

[janhilgard]

LGTM. Clean one-line fix with clear regression tests.

The logic is correct: when M <= max_rotating_size, no eviction can happen yet, so forcing the full tail back in just collapses sparse prefill into dense work. The M > max_rotating_size guard is the right condition.

The Nemotron 64K benchmarks speak for themselves — 2x speedup over dense after the fix, vs being slower than dense before.