- GC test: cutoff is in the future, so every metric is stale by timestamp
(not "nothing is stale"); only live handles retain.
- GC loop: drop the inaccurate "no Instant alloc" rationale.
- bench: the handle path no longer calls Instant::now() (it uses fastant);
reword the HANDLE_TOTAL_WORK rationale.
- rename test_unsafe_counter -> test_counter to match the type rename.
The counter is now a plain atomic (no longer UnsafeCell, no longer
sharded), so the "Unsafe" prefix is a misleading leftover. Rename to
Counter. Also fix a stale bench comment that claimed the ShardedAtomic
variant mirrors production.
Adds benches/counter_contention.rs comparing the production CounterHandle
against reconstructed baselines (pre-optimization single-atomic + CAS +
Mutex<Instant>) under tokio-task contention, plus counter-only variants.
Uses the real stats_manager types so the numbers reflect shipped code.
Placed between the sharding and fastant commits so each can be benchmarked
independently: at this commit prod = sharded counter + Mutex<Instant> touch.
fanyang