v0.2.0

Added

• /spark profiler start --alloc on Windows, producing an uploaded spark
  ALLOCATION profile weighted by sampled UCRT allocation bytes.
• /spark profiler start --alloc on Linux x86-64 using atomic BDS ELF import-slot
  hooks for glibc allocation entry points.
• /spark profiler start --alloc-live-only for retained-allocation call trees and
  leak-candidate analysis by allocation stack and age.
• Byte-based allocation stack sampling with spark's default 524287-byte interval,
  a randomized systematic sampling phase, a preallocated hook event pool,
  deferred module/symbol resolution, per-second windows, and
  --only-ticks-over support.
• CMake FetchContent integration for funchook v1.1.3, avoiding an unavailable
  Conan Center package.

Changed

• Reject ambiguous export-only UCRT symbol names during Windows symbolization and
  fall back to module+0xRVA instead of displaying unrelated nearest exports.
• Export allocation sample count, dropped sample count, estimated sampled byte
  weight, observed request bytes, interval, backend, and coverage through spark
  extra_platform_metadata.
• Report every candidate allocation entry point as active, aliased, missing, or
  unhookable in profiler status and exported platform metadata.
• Follow sampled allocations through realloc and free entry points across threads,
  reporting live/freed estimates and sampled allocation lifetimes on both native
  backends.
• Preserve the process's effective Linux allocator implementation, including
  LD_PRELOAD interposition, while redirecting BDS import slots.
• Harden native hook lifecycle handling: prepare and retain trampolines separately
  from entry-hook installation, retain disabled entry hooks between sessions,
  stabilize and verify thread suspension, retry thread restoration, and require
  complete entry-hook removal plus hook/trampoline quiescence before plugin unload.
  An unrecoverable cleanup failure terminates the process instead of pinning old
  plugin code or permitting an unsafe reload.
• Treat the allocation aggregator as a fallible service: failures disable capture,
  block partial-profile export, remain visible through profiler status, and allow
  a clean subsequent session after stop/cancel.
• Report failed allocation sessions explicitly, discard their incomplete data on
  stop/cancel/timeout, and confirm when the backend is ready to start again.
• Write valid gzip-compressed local profiles atomically and automatically save a
  local copy when upload fails, so completed captures are not lost to the network.
• Replace the finite repeated threshold table with constant-time systematic byte
  sampling using a fresh uniformly random phase per session. Allocation weights
  use successful requested sizes and every sampling point is attributed to the
  allocation that contains it.
• Expand native coverage to recalloc, aligned allocation families, internal UCRT
  base exports, and direct HeapAlloc/HeapReAlloc, while avoiding duplicate hooks
  for alias exports.
• Encapsulate Windows allocation backend state in the AllocationSampler instance
  instead of file-level mutable globals.
• Removed the temporary /spark alloc start|info|stop|reset|test validation
  commands and their allocation/free counter implementation.
• Stop sampling before collecting plugin/world export metadata so the profiler's
  own metadata allocations are not included in allocation profiles.

Limitations

• Native allocation profiling, sampled lifetimes, and retained/live analysis cover
  the selected server thread on Windows and Linux x86-64.
• Custom thread selection, direct virtual-memory calls, and custom allocator
  activity below its backing Windows heap allocation are not yet supported.

Full Changelog: v0.1.1...v0.2.0