dashboard: sync PR #284/#285 results and refresh Qwen3.6 frontier

[skyrocket2026]

Summary

• Adds the missing PR perf(qwen3.6): int8 KV + tensor-core flash-decode for the hd256 full-attn layers (long-context) #284 and Opt/gdn conv l2norm #285 dashboard entries (bot ran these evals on feat/polaris-receipts before the branch fully synced to main).
• Refreshes the Qwen3.6 per-context frontier chart to PR perf(qwen3.6): int8 KV + tensor-core flash-decode for the hd256 full-attn layers (long-context) #284's actual measured numbers now that it's merged: 16k 338.55→367.97 tok/s, 32k 301.19→345.70 tok/s (413.09/405.37/389.10 at 128/512/4k).
• Adds PR perf(qwen3.6): int8 KV + tensor-core flash-decode for the hd256 full-attn layers (long-context) #284 to landed_qwen36.

All other eval-bot fixes from the feat/polaris-receipts branch (16k/32k baseline fallback, sweep condition, vastai PATH, instance default, log repo URL) turned out to already be present on main independently — cherry-picks came back empty, confirmed via direct content diff.

Test plan

• Validated both dashboard/data.json and dashboard/data.js parse as valid JSON
• Confirmed data.js is regenerated from data.json via the same logic as eval/pr_eval_bot.py's write_dash()
• Diffed PR perf(qwen3.6): fused router GEMV + bitonic top-k (grid-completion, decode) #282/perf(qwen3.6): fuse Q-gate + O-proj Q8_1 quantize into one launch #283 entries between branch and main — identical, confirming safe merge