dashboard: backfill PRs #455 #463 #464 #465 from latest evals

[skyrocket2026]

Summary

• PR perf(qwen35): windowed prefill attention for Qwythos long context #455 → XL @ 64k prefill (4394.9 pp tok/s, 0455-ef6bc69) + journey merge
• PR perf(qwen35): int8 tensor-core prefill attention for Qwythos #465 → XL prefill attention (7828.06 pp tok/s @ scored ctx, 0465-0217e1f)
• PR perf(qwen35): fp16 shared-memory tiles for windowed prefill attention (+3–5% pp) #463 → M fp16 smem tiles (5509.41 pp tok/s, 0463-ecc22ff)
• PR perf(qwen35): fused Q4K/Q6K->int8 dequant + token-parallel GDN conv + lane-parallel fallback attention (+78% pp @32k on top of #465) #464 → XL @ 32k prefill (14057.87 pp tok/s, +83.5%, 0464-b699eb5)
• Repair Qwen3.5 optimization journey (landed_qwen35)

Test plan

• Backfilled from sparkinfer-log run IDs
• Dashboard shows new PR rows and prefill frontier after merge