Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Diagnosis of the missing page retention hysteresis in UintptrFree. Also
drops the now-stale "untracked" qualifier from CLAUDE.md.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

UintptrFree now keeps a drained shared page as the carve target of its size
class instead of returning it to the OS at once. A workload that repeatedly
drains a size class no longer pays a mmap/munmap round trip per turnover: an
alloc/free loop at 64 B or 4 KiB drops from 200,024 mmaps and 600,000 munmaps
to a single allocator page. Retention is bounded by one empty page per class.

Trim hands the retained pages back. Tests and benchmarks call it before
asserting the counters and the mapping set are zero, so those assertions keep
their full strength; asserting after Close instead would be vacuous, as Close
ends with *a = Allocator{}.

Pages backing a single big allocation are unmapped immediately, as before.

See HANDOFF.md for the diagnosis and the measurements.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>