Summary

Recurrent backup cells are expanded for speculative decoding, but the server does not shrink them back before later prompt-cache/checkpoint restore. That means checkpoint restore is happening against a different recurrent-cache topology than the one used when the checkpoint was made.

Fixes #49.

Code Changes

• Add a server prefill shrink step before prompt-cache save/load so recurrent memory returns to the non-speculative topology once draft backup cells are no longer active.
• Remove speculative backup sequences before shrinking recurrent memory.
• Skip shrinking while a slot is processing or still owns a draft backup.
• Sanitize recurrent cell metadata after resize by dropping invalid sequence ids, clearing stale source rows, rebuilding tail pointers, and clamping cached range metadata.
• Guard recurrent source-row lookup so invalid metadata falls back to the zero state instead of reaching backend GET_ROWS with an invalid row.

Validation

• Built llama-server in a Windows CUDA Release build.
• Checked the branch diff to confirm only the recurrent-cache fix files are included.