[GoDeskio]

Summary

• Adds optional SSD tiering (tiering-enabled, off by default). Keys stay in RAM. String values larger than tiering-min-value that have been idle for tiering-idle-seconds are appended to tiering-path and replaced by a checksummed pointer.
• A read promotes the value back into RAM. Compaction rewrites live records and renames the new log into place, so a crash before rename keeps the previous file. A torn tail is dropped the next time the log is opened.
• RDB save and AOF rewrite read the log, so a restart still has the bytes. INFO reports spill, promotion, and dead-byte counters.
• The value log layout is original. The shape (append-only value log plus a pointer, keys left in RAM) follows the public description of memcached extstore and KeyDB FLASH, both BSD. Redis on Flash was not used.

Tests

• ./runtest --single unit/tiering (9 tests): spill and promote, delete, expire, mixed read/write/expire, RDB reload, crash before compaction rename, torn tail, AOF rewrite, and a dataset larger than 5× tiering-max-ram spilling back under the cap. All passed.
• Full suite on this branch: 5876 passed, 2 failed. One failure was this new compaction test (the shared log was not yet half dead, so compaction did not run). After moving that case onto a fresh log it passes. The other was integration/repl-fullsync-compression ("Failed to start fake primary") while another suite was running; rerun alone: 23 passed, 0 failed.