feat(prefill): add tiered KV hot promotion and offload

[FluffyAIcode]

Summary

• promote every successful remote snapshot into Primary's bounded hot LRU for repeat decode workloads
• retain the original snapshot in allens cold/offload cache; Primary eviction never deletes the offload copy
• add thread-safe cache budget resize with cold LRU eviction and lease safety
• adapt allens cache capacity from physical memory minus active MLX model bytes and a 2GiB reserve, bounded by 0.25GiB minimum and 8GiB ceiling
• label kv:// mounts as hot or cold-offload and expose promotion metrics through existing prefill telemetry

Tests

• platform-neutral Python suite — 837 passed, 1 skipped
• TypeScript SDK — 54 passed, 100% coverage and typecheck
• remote import promotion, promotion failure isolation, adaptive budget bounds, cache resize/lease behavior, mount tiering, installer and launchd tests pass
• Python compile, shell syntax, plist lint and diff checks pass

Live validation after merge

Redeploy allens, run one strict remote prefill, replay the exact prefix to prove a Primary local hot hit with no new remote job, then pressure the Primary LRU and verify the allens cold copy remains fetchable.

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.