fix(prefill): preserve unrelated snapshots during jobs

[FluffyAIcode]

Summary

• reserve the configured conservative final-snapshot estimate instead of the worker's entire cache budget
• retain atomic final publication and adaptive-shrink protection from PR fix(prefill): atomically reserve and publish snapshots #187
• preserve unrelated Generator and Critic snapshots across subsequent Prefill jobs
• add regression coverage proving old and new content-addressed snapshots coexist

Root cause

The first real full-context run succeeded after PR #187, but its full-budget reservation evicted every existing snapshot at each new job. A later Generator miss therefore erased the prior Critic snapshot, preventing cold restore despite correct atomic publication.

For the measured 1,010-token Critic, the configured reservation is about 404 MB while the actual imported final snapshot is roughly 90 MB. The 1+ GiB cache budget safely accommodates the estimate and unrelated snapshots.

Tests

• precise reservation and unrelated snapshot preservation — 27 targeted tests pass
• full platform-neutral suite — 867 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.