Phase 11.8: batched fast_path counter updates

[jayakasadev]

Summary

Phase 11.8 of the SNMALLOC_STATS overhead-reduction series. Moves the
per-alloc ++stats.fast_path_allocs store out of the small-alloc fast
path into a single batched pre-credit at slab-refill time.

The slow-path (small_refill / small_refill_slow) already runs once
per slab refill; it now also knows the refill count N (number of
objects transferred from the freshly-popped slab into
fast_free_list) and credits stats.fast_path_allocs += N once,
letting the fast path skip the per-alloc store entirely.

The refill count is plumbed back from
FrontendSlabMetadata::alloc_free_list via a new uint16_t& out
parameter, computed as
sizeclass_to_slab_object_count(sizeclass) - remaining:

• Exact for freshly-built slabs (where alloc_new_list loaded the
  builder with slab_object_count objects).
• Upper-bound (bounded by the slab object count, at most ~256 for
  the smallest sizeclasses) for slabs recycled from the per-sizeclass
  stash.

The counter may briefly read ahead of real consumption by at most one
slab worth between refills — acceptable for observability.

Before / after (Apple M4 Pro, fat-LTO, 5 runs per variant)

Group	11.6 ratio	11.8 ratio	acceptance (<=1.02)
small_allocs	1.0774	1.0155	PASS
medium_allocs	1.0398	1.0202	FAIL*
mixed	1.0310	1.0290	FAIL

* medium_allocs is within bench noise on this host (per-run-pair
median 0.9941, i.e. statistically indistinguishable from OFF once a
run-1 cold-cache outlier on the OFF side is accounted for).

A 3-run replication on a separate invocation reproduced the same
shape: small ~1.018, medium ~1.015, mixed ~1.026.

Acceptance verdict

PARTIAL.

• small_allocs (the targeted group, where the per-alloc fast-path
  counter dominated the iteration mean) passes the strict <=1.02 spec
  cleanly at 1.0155, a ~80% reduction of the previous 1.0774
  over-budget portion.
• medium_allocs lands at 1.0202 with the per-run-pair median in
  favour of the BASIC build.
• mixed (1.0290) still misses the strict 1.02 spec. It blends
  large-class paths that do not benefit from the small-class batching
  done here, and still pays the symmetric per-dealloc
  fast_path_deallocs store on the dealloc hot path.

Phase 11.9 is filed as a follow-up to apply the same
single-combined-counter approach to the dealloc-side counters.

Build / test status

• cmake -B build -DSNMALLOC_STATS_BASIC=ON && cmake --build build -j4 clean.
• ctest -R "fast_path_counters|statistics" 4/4 pass.
• cargo test --features stats-basic in snmalloc-rs/: full suite green.

Files touched

• src/snmalloc/mem/corealloc.h — remove per-alloc store; add batched
  pre-credit at the two refill sites.
• src/snmalloc/mem/metadata.h — alloc_free_list reports refill count.
• docs/heap-profiling-benchmarks.md — Phase 11.8 section with full
  5-run tables, acceptance verdict, and reproducer.

Test plan

• Build with -DSNMALLOC_STATS_BASIC=ON passes
• ctest -R fast_path_counters and -R statistics pass
• cargo test --features stats-basic passes
• cargo bench --features stats-basic --bench stats_bench ran 3+ times
• cargo bench --bench stats_bench (OFF baseline) ran 3+ times