fix(prefill): cap reservation at sliding window

[FluffyAIcode]

Summary

• estimate retained snapshot bytes with min(prompt_tokens, sink + window)
• apply the same estimate to upfront capacity validation and active reservation
• preserve complete prompt processing and full-context Critic semantics
• prevent long auto-loop histories from falsely exceeding cache capacity after sliding-window saturation

Root cause

The failed 2,731-token turn was estimated as 1,092,400,000 bytes, 18.66 MB over the 1 GiB cache. MLX retains only sink 4 + window 2,048 = 2,052 KV tokens, so the correct conservative estimate is 820,800,000 bytes and fits safely.

Tests

• exact 2,731-token production regression
• below-window and no-window estimate behavior
• upfront validation and reservation paths
• targeted suite — 32 passed
• full platform-neutral suite — 881 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.