Heap corruption from unsafeForcedSync in diarizer pipeline (iOS 26.4)

[EcoeOS]

Environment

• iOS 26.4.1 (Build 23E254)
• Device: arm64, chip=V57AP, cores=6/6, mem=7.49 GB
• FluidAudio commit: 044bb0b
• Swift 6 strict concurrency enabled

Bug

During live diarization sessions using SpeakerGate (with a saved voice print loaded via DiarizerModels + EmbeddingExtractor), we observe repeated warnings followed by an eventual heap corruption crash:

Warnings (every audio buffer)

Potential Structural Swift Concurrency Issue: unsafeForcedSync called from Swift Concurrent context.

Crash (intermittent, on FluidAudio's internal serial queue)

Thread 40: EXC_BREAKPOINT (code=1, subcode=0x...)
libsystem_malloc.dylib`_xzm_xzone_malloc_from_freelist_chunk.cold.1:
  "BUG IN CLIENT OF LIBMALLOC: memory corruption of free block"

The crash is on a FluidAudio-internal serial queue (not our code's queue). The unsafeForcedSync warnings suggest a Swift Concurrency data race inside FluidAudio that corrupts the heap over time.

Reproduction

1. Initialize DiarizerModels.download() + SpeakerManager with a known speaker
2. Feed 48kHz audio buffers from AVAudioEngine.inputNode.installTap
3. Run speaker verification via extractSpeakerEmbedding every ~1.5s
4. Crash occurs after 1-5 minutes of continuous operation

What we've ruled out

• Our audio buffers are copied before dispatch (D-015 pattern)
• SpeakerGate feeds resampled [Float] arrays, not raw AVAudioPCMBuffer
• No concurrent access from our code — all FluidAudio calls are serialized through gateQueue and diarizeQueue
• The unsafeForcedSync warnings originate from FluidAudio internals, not our call sites

Expected behavior

No heap corruption during sustained diarization sessions.