fix(eval): Polaris TDX receipts on real GPU evals

[skyrocket2026]

Summary

• Fix Polaris TDX end-to-end on SSH GPU evals: correct API base URL, read verification from top-level response, store stdout/quote fields in receipts, and fix TDX hash + verify.py pubkey checks
• Auto-pin Qwen3.6 model SHA in single-model evaluate.sh runs; forward RESULT_JSON / POLARIS_ATTESTATION before vast_eval stdout tail so the bot can attest
• Add helper scripts: run_bot.sh, run_polaris_smoke.sh (~20s TDX smoke), run_polaris_test.sh (dual eval + TDX)
• Dashboard: record PR perf(qwen3.6): fuse decode kernel launches — shared-expert + GDN conv/L2-norm #239 eval result (eval:none, 419.24 tok/s @512)

Test plan

• ./eval/run_polaris_smoke.sh — TDX attest + intel_verified=True
• Qwen3.6-only PR perf(qwen3.6): fuse Q-gate + O-proj Q8_1 quantize into one launch #283 SSH eval → POLARIS_ATTESTATION → receipt → verify.py passes DCAP checks
• ./eval/run_bot.sh --polaris --only-pr <N> --reeval on SSH box (production bot path)

Made with Cursor