docs: replace pad evidence with valid paired run

[shinaoka]

Summary

• reclassify the earlier 7.5%-load pad candidate as INCONCLUSIVE
• record a fresh same-domain baseline/candidate pair satisfying the frozen <2%/<20% load gate
• publish corrected medium/large timings and control evidence

Part of #240 and #213.

Corrected medium speedups

• serial rank 2/4/8: 2.46x / 4.42x / 7.65x
• 4T rank 2/4/8: 1.34x / 2.09x / 3.42x
• negative/non-unit serial: 2.29x / 1.99x
• max control regression: 8.2% (<10%)

Exact read-only reviewer-flash review: Correct-to-merge.