fix(prefill): atomically reserve and publish snapshots

[FluffyAIcode]

Summary

• preflight estimated final-snapshot capacity before enqueuing model work
• reserve the worker's full current cache budget before each single-concurrency MLX job
• prevent adaptive cache resize from consuming active reservations
• atomically publish and lease the final snapshot before optionally retaining intermediate boundaries
• prioritize final snapshot correctness over intermediate cache density
• raise allens adaptive cache floor from 0.25 GiB to 1 GiB and reduce reserve from 2 GiB to 0.5 GiB
• bind Primary gRPC to 0.0.0.0:51051 so allens can reach its advertised Thunderbolt address

Root cause

The failed full-context Critic completed roughly 1,200 seconds of model compute, then published snapshots one by one into a 256 MiB adaptive LRU. Concurrent adaptive shrink and 66 evictions (1.07 GiB evicted) removed the final snapshot before the subsequent lookup/lease, producing computed snapshot was not discoverable.

Correctness guarantees

• capacity rejection occurs before model compute
• active reservations survive adaptive budget refresh
• final snapshot publication and lease creation share one store lock
• no local Prefill or semantic fallback is introduced

Tests

• reservation, adaptive shrink, upfront rejection, cancellation, atomic final pinning, and remote orchestration coverage
• targeted suite — 27 passed
• full platform-neutral suite — 863 passed, 1 skipped
• Python, shell, plist, 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.