Dense 27B sparse V validation (clean runs, M5 Max 128GB)
Model: Qwen3.5-27B Q8_0 (dense)
Date: 2026-03-27
All stale processes killed, no GPU contention.

=== A/B/A SHORT CONTEXT (order swap to rule out thermal) ===
Run 1 (WITHOUT): 16.87 ± 0.06 tok/s
Run 2 (WITH):    16.95 ± 0.04 tok/s
Run 3 (WITHOUT): 16.56 ± 0.55 tok/s
Delta: +0.5% (within noise, no regression)

=== LONG CONTEXT DECODE ===
--- WITH sparse V ---
tg128:          16.73 ± 0.07
pp8192+tg128:   298.27 ± 1.13
pp16384+tg128:  316.98 ± 0.41

--- WITHOUT sparse V ---
tg128:          16.61 ± 0.17
pp8192+tg128:   294.52 ± 4.30
pp16384+tg128:  311.24 ± 4.57

=== DELTA ===
Short (tg128):     +0.7%
8K (pp8192):       +1.3%
16K (pp16384):     +1.8%

=== CONCLUSION ===
No regression. Neutral-to-slight-positive. Trend improves with context.
Safe to enable by default for dense models.

Note: earlier run showed -5.4% regression but was contaminated by 4 stale
llama-server processes eating ~85% memory. A/B/A test confirmed no regression.
