Things we believed and then proved wrong

• Prefix caching causes 84 additional GPU kernels.FalseAn entry-point migration
• A single-entry-point uprobe gives complete launch coverage.FalseOn this stack
• CUDA Graphs alone caused the 36.1% reduction.UnprovenCompile confound. Host launches did fall 1,146 → 134.
• SETCTAID.X is a no-op.FalseThe observer stored a saved value twice
• Our 256-token prefill cap was being broken.FalseThe check was wrong
• A 4,096 µs timeslice gives longer turns than 512 µs.FalseMeasured ratio 1.0003
• Interleave HIGH gets a bigger share than LOW.False1.0007 against a 1.5 threshold