[sharooms]

Fixes #491.

RtApiAsio::callbackEvent and RtApiCore::callbackEvent invoked the user callback before copying the driver's input into stream_.userBuffer[1], so the callback always received the previous cycle's input: one buffer of extra round-trip latency that getStreamLatency() does not report. The issue has the measurements (macOS/CoreAudio and Windows/ASIO, cross-checked with RTL Utility) and how it came up; the analysis and this patch were worked out with Claude and tested on real hardware.

The change

In both functions the block that copies the input into stream_.userBuffer[1] moves in front of the user callback. The block itself is unchanged except for its condition, which now carries the existing "don't bother draining input" rule (handle->drainCounter == 0) instead of a goto unlock after the output. The drain counter is still advanced after the output, exactly as before.

• ASIO: at bufferSwitch(index) the input half index has just been filled by the driver, so it can be read before the callback. Duplex and input-only streams both get the current input.
• CoreAudio, one device (duplex or input-only): the callback now sees the input of the same IOProc call.
• CoreAudio, two devices: unchanged. The input device's IOProc never invoked the callback and still only stores the input; the output device's IOProc never read input.
• stream_.deviceBuffer is shared by the input and output conversion. They still run one after the other within one call, only in the other order.
• Side effect: after a restart, a one-device CoreAudio duplex stream no longer hands its first callback the last input of the previous run.

WASAPI, JACK, ALSA, PulseAudio, OSS and DirectSound are not touched.

The open #411 also edits the CoreAudio input copy block that this change moves. Whichever of the two lands second needs a mechanical rebase; I am happy to do it on this side.

Tested

• macOS, CoreAudio: Audient iD4 (duplex with a cable loopback, and output-only + input-only streams), MacBook speakers + microphone (two devices). Round trip one buffer shorter at 64 / 256 / 1024, every start identical, no xruns; the two-device path measures as before.
• Windows, ASIO: RME Digiface USB with an external ADAT converter. Same result. Inside a DAW, buffer-size changes in the driver panel (ASIO reset), unplugging the interface, sample-rate changes and repeated start/stop ran through the patched callback and behaved as with the unpatched build.
• Builds: Apple clang on macOS, clang-cl 20 with the MSVC 14.50 headers on Windows (ASIO + WASAPI).