[kvlonge]

Why

The mimalloc allocator from #889 (v3) keeps freed memory for up to 1,000 ms before returning it to the OS. On the 3,000-model inspection benchmark, sqb plan --json peak memory rose from about 890 MB to about 1,160 MB (1.26x) in every run. That exceeds the 1.25x release comparison limit and blocks the next release.

Changes

• The CLI console-script entry point sets MIMALLOC_PURGE_DELAY=10 (mimalloc v2's default) before it imports the CLI, which loads the native extension. A value already set by the user is kept (os.environ.setdefault).
• The Rust side cannot set the option because the workspace forbids unsafe, and the root package may not run code at import time.

Measured on main, plan peak RSS / compile --no-cache wall time:

• 1,000 ms (current default): 1,094-1,165 MB / 8.0-8.1 s
• 10 ms (this change): 950-1,017 MB / 8.2-8.4 s
• 0 ms: ~906 MB / ~8.8 s
• 0.126.5 (before mimalloc): 887-950 MB / ~10 s

Verification

• New fresh-process integration test: importing the entry module does not load sqlbuild._native; running it sets the 10 ms default when unset and preserves a user value. Both cases fail without the fix.
• make check-ci passes.
• The release PR's performance comparison re-runs after merge.