…te and fall back for older Windows and Xbox similar to Posix other than Linux and Android.

…the buffers transmitted during the final shutdown flush... add them to the test assertion

…fferentiating based on what the non-existent condvar meant previously

…meaningless than before...

we are not only bounded by the last flush, but since we now trigger with almost immediate wake on essentially any count in the buffer... we can get requests as high as the number of items we send. The important part is the total count and that one has been stable throughout.

… could still be running after the example process terminated due to the crash

…unting (#1556)

* test(logs): add stress test for batcher reinit hang

Spawns 8 producer threads continuously logging while the main thread
does 10 cycles of sentry_init/sentry_close. Reproduces a condvar
corruption hang in the batcher under TSan.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): replace static global with dynamic allocation and refcounting

The static g_batcher persisted across init/close cycles, causing condvar
corruption when sentry__batcher_startup re-initialized the condvar while
old threads still used it. Dynamic allocation with refcounting solves
both the condvar corruption and the lifetime problem.

Introduces sentry_batcher_ref_t with acquire/release/swap API that
encapsulates a spinlock to make concurrent access from producer threads
and shutdown safe.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): release old batcher on swap in startup

When the previous startup failed to spawn its thread, shutdown_begin
returns false and shutdown_wait is never called. The old batcher stays
in g_batcher and leaks when startup swaps it out without releasing.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): NULL-guard batcher in shutdown_wait

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Update CHANGELOG.md

* fix(batcher): drain buffer items in release to prevent leaks

A producer that acquired a ref before shutdown could enqueue items after
the final flush. These items were leaked when the batcher was freed.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(batcher): use SENTRY_MAKE macro for allocation

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): destroy condvar before freeing batcher memory

Add sentry__cond_destroy for all platforms (pthread_cond_destroy on
POSIX, CloseHandle on pre-Vista Windows, no-op on Vista+) and call it
in sentry__batcher_release. Without this, each SDK re-initialization
leaks kernel handles on pre-Vista Windows and violates POSIX which
requires pthread_cond_destroy before freeing memory containing a
pthread_cond_t.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Revert "fix(batcher): destroy condvar before freeing batcher memory"

This reverts commit 467bfb5ecb2c4528e1048dab1c467203fdc29cdf.

* refactor(batcher): merge two-phase shutdown into single function

The two-phase shutdown (begin/wait) was designed to signal both the logs
and metrics threads in parallel before joining either. With #1558
replacing the condvar with a level-triggered waitable flag, thread wake
latency drops to sub-millisecond (futex/WaitOnAddress), making the
parallel signaling unnecessary.

Merging into a single function also fixes a race where a concurrent
sentry_init() could replace g_batcher between the signal and join steps,
causing the new batcher to be shut down instead of the old one.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test(metrics): add reinit stress test

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test: speed up reinit stress tests

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): use lock-free peek in crash-safe flush path

The crash-safe flush functions called sentry__batcher_acquire, which
spins on a spinlock. If the crash occurs on the thread holding that
lock, the signal handler deadlocks. Replace with a new lock-free
sentry__batcher_peek that reads ref->ptr via atomic load without
taking the spinlock or bumping the refcount (safe because the process
is dying).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): use pointer-width atomics in batcher_peek for Windows

On Windows x64, `long` is 32-bit but pointers are 64-bit. Use
InterlockedCompareExchangePointer on Windows and __atomic_load
elsewhere to avoid truncating the pointer.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs(batcher): add header comments for ref management functions

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(batcher): fix race between force flush and reinit

force_flush_begin and force_flush_wait each acquired the batcher
independently. A concurrent sentry_init could swap the batcher between
the two calls, causing wait to flush the new empty batcher while the
original data is lost.

Fix by returning an opaque token (the acquired batcher ref) from begin
and passing it to wait, ensuring both operate on the same instance.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Update src/sentry_logs.c

Co-authored-by: Mischan Toosarani-Hausberger <mischan@abovevacant.com>

* Update sentry__metrics_startup

* docs(batcher): clarify ownership semantics of options refs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Mischan Toosarani-Hausberger <mischan@abovevacant.com>