perf(prefill): export only final MLX snapshot

[FluffyAIcode]

Summary

• keep full MLX Prefill computation over every input token
• export, KakeyaLattice-encode, and zlib-compress only the final snapshot
• accept final-only worker results while retaining compatibility with engines that publish every boundary
• preserve atomic capacity reservation and final lease guarantees
• add tests proving one export for multi-block prompts and final-only remote job discovery

Root cause

The worker previously exported a growing full KV snapshot every 64 tokens. A 1,569-token Critic context therefore performed roughly 25 full exports and compression passes. Since each later snapshot contains a longer prefix, serialization work grew approximately quadratically and made distributed Prefill unusable compared with single-node Prefill.

The final chained hash commits all preceding blocks, so intermediate full snapshots are not required for correctness.

Tests

• final-only MLX export, cache publication, lease, and remote worker paths — 30 targeted tests pass
• full platform-neutral suite — 870 passed, 1 skipped
• Python compile 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.