- Use functional style (map + match) in collect_batched_heap_live_samples
- Only create ProfileIndex when heap-live tracking is enabled
- Replace 32 repetitive I/O profiling lines with a loop
- Use filter_map in sample type filter method
- Add early bail-out in free_allocation when heap-live is disabled

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Replace default SipHash with a simple bit-mixing hasher optimized for
pointer addresses. Since pointers are already well-distributed, we use
`ptr ^ (ptr >> 4)` instead of expensive cryptographic hashing.

This reduces overhead in untrack_allocation() which is called on every free.

Co-Authored-By: Claude Opus 4.5 <[email protected]>

- Cargo.lock: drop unrelated lockfile churn so the diff vs master
  only contains entries cargo actually pulled in for dashmap.
- SampleData / LiveHeapSample: Arc-share frames and labels so the
  in-flight sample message and the heap-live tracker share one
  allocation each, and re-emitting tracked entries on every export
  becomes Arc::clone instead of deep clone.
- realloc: guard against NULL return so we don't free-track a still-
  live prev_ptr.
- visibility: tighten heap-live helpers to pub(crate); drop the
  unused clear_live_heap_tracker / live_heap_tracker_len helpers.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

ProfileIndex is effectively a singleton per request (sample_types fixed at
startup, tags resolved once per rinit), but it was being deep-cloned on
every prepare_sample_message, every HashMap insert, every heap-live track,
and every heap-live emission.

Wrap it in Arc so:
 - SampleMessage::key and LiveHeapSample::key share one allocation.
 - HashMap<Arc<ProfileIndex>, InternalProfile> keys via Arc's Hash/Eq
   (which delegate to ProfileIndex), so behavior is unchanged.
 - The heap-live re-emission path becomes a pure Arc::clone instead of a
   deep Vec<ValueType> clone.

Behavior-preserving refactor. Pairs with the existing Arc<Backtrace> /
Arc<Vec<Label>> work on SampleData.

The previous implementation paid for a Profiler::get() (OnceLock atomic
load) + AtomicPtr load + SystemSettings deref on every efree/erealloc,
even when heap-live profiling was disabled. That overhead is likely
behind the timeline-memory bench regression observed in the PoC, since
that benchmark exercises the free path heavily without enabling
heap-live.

Add a static AtomicBool that mirrors profiling_heap_live_enabled and
gate the alloc handlers (alloc_prof_free, alloc_prof_realloc) on it
*before* touching Profiler::get(). When heap-live is off the off-path
collapses to a single relaxed load + branch.

The static is set in Profiler::new and cleared in Profiler::stop /
Profiler::kill, so teardown is also safe (any late free during shutdown
short-circuits cleanly).

Profiler::is_heap_live_enabled now reads the same static so all gates
agree on one source of truth, dropping the AtomicPtr/deref dance on the
allocation-sampling hot path as well.

…ce lock contention

DashMap::iter() holds a per-shard read lock for each entry while in
scope. With the previous code, handle_sample_message was called while
the lock was still held, blocking concurrent efree calls on PHP threads
for the full duration of a 4096-entry export iteration.

Snapshot all values into a Vec first (cheap: just Arc::clone per field),
then process with no locks held.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…n free/update methods

Call sites in the allocation hooks already gate on HEAP_LIVE_ENABLED
before calling free_allocation and update_allocation_size. The inner
checks were dead code and misleading. Replace with debug_assert and a
precondition doc comment.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

The XOR-shift mix (ptr ^ (ptr >> 4)) leaves sequential bump-allocator
addresses correlated in the high bits, clustering them onto the same
DashMap shards under ZTS workloads. FxHasher's multiply-rotate mix
fully avalanches bits, spreading sequential pointers evenly across all
16 shards. rustc-hash is already a direct dependency.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

…eap-live sample values

Replace fragile string matching on "heap-live-samples"/"heap-live-size"
in collect_batched_heap_live_samples with SampleTypeFilter::filter(). A
rename or typo in the type strings previously produced silent all-zero
heap-live exports. Derive Clone on SampleTypeFilter and pass a clone to
TimeCollector.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <[email protected]>

Previously, in-place reallocs called update_allocation_size() which
mutated the tracked allocation_size from S0 to S1 in place. This broke
the upscaling invariant: the Poisson estimator requires that the
sampling decision was made at the size used in the upscale formula, but
the allocation was sampled at S0 while the upscaler saw S1.

Fix by treating in-place reallocs the same as pointer-moving reallocs:
untrack the old allocation and re-sample at the new size. The sampling
decision is then always made at the reported size, keeping the estimator
unbiased. In-place reallocs are rare in PHP, so the performance impact
is negligible.

Remove update_allocation_size() from Profiler, allocation/mod.rs, and
both allocation handler files entirely.