[SP] Audio: fix crackle in cinematics, better resampling and limiting, lower latency

[Rugaliz]

Summary

Fixes crackling audio during cinematics (heard in the opening logos) and improves SP sound quality and latency.

Changes

• Cinematic crackle: cinematic audio and streamed music both append to the same raw sample ring. The main menu music is started by a console command while the opening logos video is still playing, so the two interleaved, and the first audio packet of a roq video (~0.5 s, 23520 samples at 44.1 kHz) was larger than the whole 16384-sample ring. The ring is now 65536 samples, music buffering keeps its original 16384 look-ahead (MUSIC_RAW_LOOKAHEAD), and music is not streamed while a cinematic with a soundtrack is playing (CIN_IsPlayingAudio). Frame rate was ruled out as a cause (tested caps of 30/125/250 and vsync).
• Resampling: ResampleSfx used nearest-neighbour conversion, which mirrors high frequencies into the audible range (harsh, gritty 11/22 kHz sounds). New s_quality cvar: 0 linear/averaging, 1 cubic, 2 windowed-sinc (default). Costs ~0.35 s extra per map load (233 sounds: 373 ms vs 20 ms at s_quality 0), nothing during play. Cinematic audio is linearly interpolated instead of repeating samples. s_quality 0 is the cheap option for slow machines.
• Limiter: hard clipping of the mix is replaced by a soft limiter that only acts above 0.75 of full scale (quieter sound is untouched). With developer 1 the log reports when the mix exceeds full scale.
• Latency: the SDL callback buffer defaults to ~11 ms instead of ~23 ms (the mix ring keeps its previous size, s_sdlDevSamps still overrides), and the mix pre-step adapts to the device buffer and recent frame times (s_mixPreStepAuto, default 1; s_mixPreStep stays the upper bound, default 0.05). The pre-step settled at ~30 ms at 125 fps and ~28 ms at 144 fps. s_mixPreStepAuto 0 restores the old fixed behaviour.

Not included: native 48 kHz output. Voice lines under sound/chars/ stay MP3-compressed and are decoded on the fly by mp3code, which only outputs 44.1/22.05/11.025 kHz, so that would need a separate change.

Testing

• JO SP on Linux (PipeWire): the intro logos played in full with zero ring overflow messages (previously dozens per second of video), new game cinematics and briefing, several map loads, 125/144 fps caps; no audible artifacts reported by the tester.
• Offline check on a real 22.05 kHz game sound: energy above the source Nyquist drops by ~11 dB with the new resampler (Python re-implementation of the algorithm).
• JA SP, the MP engine (shares sdl_sound.cpp) and JO SP build cleanly. JA SP and MP were not run, and Windows/macOS were not tested. The 32-bit MSVC OpenAL code path was not touched or tested.

Notes

Original code, contributed under the project's GPLv2 license (the spline, sinc and window formulas are standard math, no third-party code was copied). Written with AI assistance (Claude Code) and reviewed/tested by the submitter.

🤖 Generated with Claude Code