feat(bench): add three-phase prefill fleet benchmark

[FluffyAIcode]

Summary

• add one terminal command that safely ensures Primary/allens services, then runs remote compute, Primary hot hit, and allens cold restore phases while leaving services running
• measure append/prefill latency, TTFT, decode latency, generation ms/token, decode tok/s, E2E tok/s, hit source, transferred bytes and failure deltas
• persist privacy-safe benchmark runs, live progress, stage detail and summary history in NetworkState with 200-run retention
• add public benchmark list/live/detail/stages APIs and API-key-protected create/update endpoints
• add a kakeya.ai Benchmarks tab with live status, phase cards, history and detail tables

Privacy

Reports reject prompt text, token IDs, cache keys, block hashes, private IPs and /Users paths. Public config contains only logical model/topology settings and a salted prefix ID.

Tests

• full platform-neutral Python suite — 846 passed, 1 skipped
• new benchmark/network/distributed modules — 100% coverage
• TypeScript SDK — 54 passed, 100% coverage and typecheck
• schema derivation, gate failures, persistence/retention, API auth/pagination, redaction and dashboard surface tests pass
• shell syntax, Python compile and diff checks pass

Live validation after merge

Deploy to Primary, run bash scripts/run_prefill_architecture_benchmark.sh --output-tokens 32, verify all three phases, then confirm local and public /v1/network/benchmarks plus the kakeya.ai Benchmarks tab.

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.