[JasperGeurtz]

#90

• Reduce async FrameBuffer memory from 33.0 MB to 4.08 MB per slot (-88%, saving ~290 MB for 10 slots) (Reduce Async FrameBuffer memory usage #90):
  • Route static map/region reads (groundHeight, isWalkable, isBuildable, regions) directly to the main shared memory buffer via AsyncClientData.
  • Omit client-to-server output buffers (strings, shapes, commands, unitCommands) from FrameBuffer (flushed directly to liveGameData via SideEffectQueue).
  • Pack the remaining 3 dynamic ranges into PackedWrappedBuffer (4.08 MB) using pre-shifted slice2 and slice3 buffer views so generated AsyncClientData methods use compile-time BWAPI offsets for both unpacked liveData and packed FrameBuffer slots with zero runtime branches.
• Only copy active units, events, and eventStrings per frame:
  • Track maxUnitCount (via initialUnitCount and UnitDiscover events) and skip copying the unused tails of units[10000], events[10000], and eventStrings[1000].
  • Reduces per-frame copy size from 4.80 MB (and 330 MB on Frame 0) to ~450 KB at game start, ~685 KB in mid-game, and ~1.19 MB in late-game (~0.05–0.20 ms on normal hardware).
• Make asyncUnsafe thread-safe on on-time frames:
  • Defer swapping the bot from liveData onto frameBuffer.peek() until the frame actually times out (remainingNanos <= 0) and control is about to return to StarCraft, avoiding mid-frame buffer swaps when the bot finishes on time.
• Consolidate async locks and eliminate spurious wakeups:
  • Consolidate lockWrite, synchronized framesBuffered(), unsafeReadReadyLock, and lastBotThrowLock onto frameBuffer.lockSize (and volatile lastBotThrow), and stop signalAll() when clearing unsafeReadReady = false.
• Double-buffer SideEffectQueue:
  • Swap queue and flushQueue under the monitor and apply side effects to liveGameData outside the lock so the bot thread is not blocked on enqueue() during shared-memory flushes.