[beinan]

Summary

Large value replacements among many small cold keys can fill the client journal even when most retained bytes are obsolete: raft-engine's normal rewrite trigger counts stale entries rather than their bytes. Before fencing a batch for insufficient disk headroom, attempt one synchronized rewrite-queue reclamation using the existing reserve, measure the directory again, and retain the fence if space is still insufficient or maintenance fails.

The binding, file format, configured budgets and durable completion barrier stay unchanged. Add a pressure-rewrite counter and regressions for continuous mixed-size replacements, sync failure, and SIGKILL immediately before and after the rewritten-data sync. The fallback is synchronous and can add latency to the triggering batch.

Refs #378.

Testing

• Full repository ci-pr-helper checks: Rust formatting, workspace Clippy/tests, Python tests, Ruff and Pyright.
• Client-journal feature tests and Clippy, including the final fault-hook changes.
• The mixed-size regression completes 40 rounds without reopening; the previous implementation fenced in round 7. Recovery compares every acknowledged latest value.