perf(prefill): segment long jobs for autoresearch

[FluffyAIcode]

Summary

• decouple model compute segments from 64-token cache hash boundaries
• default to 256-token segments, below the five-minute budget at measured allens throughput
• report live tokens_computed / total, percentage, and ETA after every segment
• retain one final snapshot export and complete semantic context
• add Karpathy-style program.md / immutable prepare.py / agent-editable candidate.py
• reject autoresearch candidates that exceed five-minute segments or violate full-context, recursive protocol, topology, no-fallback, or final-only constraints
• record accepted/rejected experiment metrics in results.tsv

AutoResearch loop

The agent changes only candidate.py, deploys, clears caches, runs the fixed workload, evaluates with prepare.py, and keeps a candidate only when all hard constraints pass and cold Critic Prefill improves.

Tests

• segmented MLX compute, progress callback lifecycle, heartbeat progress/ETA, launchd configuration, and autoresearch acceptance/rejection
• targeted suite — 42 passed
• full platform-neutral suite — 873 passed, 1 skipped
• Python, shell, and diff checks pass

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.