Louder master via makeup gain + look-ahead limiter (v0.8.0)

The master stage was a flat 0.5x attenuation into a static tanh soft-clip,
leaving ~9 dB of unused headroom: even the loudest single preset with its mix
bus maxed peaked around -7 dBFS, well below other media playback.

Replace it with a generous makeup gain (1.5x, ~+9.5 dB) feeding a feed-forward,
look-ahead brickwall limiter. A short delay line lets the peak detector pull the
gain down before a loud sample reaches the output and hold it across the
look-ahead window, so the output is mathematically guaranteed to stay under the
ceiling (0.97) - loudness without clipping, and none of the tanh grit a static
clipper adds to a hot single voice. Allocation-free (results go to scratch
fields, not a Pair) and NaN-guarded.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/build.gradle.kts

@@ -22,8 +22,8 @@ android {
         // (android.media.midi) and AAudio low-latency audio we rely on.
         minSdk = 26
         targetSdk = 34
-        versionCode = 22
-        versionName = "0.7.3"
+        versionCode = 23
+        versionName = "0.8.0"
 
         // We provide our own instrumentation runner if/when tests are added.
         testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"

app/src/main/java/space/rcmd/android/sizzle/audio/AudioEngine.kt

@@ -17,6 +17,7 @@ import kotlinx.coroutines.flow.MutableStateFlow
 import kotlinx.coroutines.flow.StateFlow
 import kotlin.concurrent.thread
 import kotlin.math.abs
+import kotlin.math.max
 import kotlin.math.sin
 import kotlin.math.tanh
 
@@ -153,6 +154,24 @@ class AudioEngine(
     /** Scratch buffer: per-channel sum for one sample (audio-thread only). */
     private val channelSum = FloatArray(Pattern.TRACK_COUNT)
 
+    // ---- master brickwall limiter (feed-forward, look-ahead) ----
+    // A short delay line lets the peak detector "see" a loud sample coming and pull
+    // the gain down BEFORE it reaches the output, so the makeup gain below can be
+    // pushed hard (bringing quiet instruments up to media loudness) while the output
+    // is mathematically guaranteed to stay under LIM_CEILING — no clipping, and none
+    // of the tanh saturation a static soft-clipper would add to a hot single voice.
+    // Gain has instant attack (drops the moment a peak enters the look-ahead window),
+    // holds for the window length, then releases smoothly toward unity. Audio-thread
+    // state only.
+    private val limDelayL = FloatArray(LIM_LOOKAHEAD)
+    private val limDelayR = FloatArray(LIM_LOOKAHEAD)
+    private var limPos = 0
+    private var limGain = 1f
+    private var limHold = 0
+    // Limiter output for the current sample (avoids a per-sample Pair allocation).
+    private var limOutL = 0f
+    private var limOutR = 0f
+
     // Per-block compact active-voice lists (audio-thread only). Rebuilt once per block:
     // the per-sample mix loop iterates ONLY the voices that are actually sounding, so it
     // never calls render() on idle pool slots (the whole POLYPHONY*TRACK_COUNT pool was
@@ -594,14 +613,54 @@ class AudioEngine(
             }
             for (k in 0 until actLiveN) { val v = liveVoices[actLive[k]].render() * LIVE_GAIN; mixL += v; mixR += v }
             for (k in 0 until actASmpN) { val v = auditionSampleVoices[actASmp[k]].render() * LIVE_GAIN; mixL += v; mixR += v }
-            val oL = softClip(mixL * MASTER_GAIN)
-            val oR = softClip(mixR * MASTER_GAIN)
+            // Makeup gain (lifts instruments to media loudness) into a look-ahead
+            // brickwall limiter that guarantees the output stays under the ceiling.
+            masterLimit(mixL * MASTER_MAKEUP, mixR * MASTER_MAKEUP)
+            val oL = limOutL
+            val oR = limOutR
             out[2 * i] = oL
             out[2 * i + 1] = oR
             if (recCapturing) recordSample(proj, oL, oR)
         }
     }
 
+    /**
+     * Feed-forward, look-ahead brickwall limiter for the master bus. Results land in
+     * [limOutL]/[limOutR] (no per-sample allocation). Runs on the audio thread.
+     *
+     * How it guarantees "no clipping": the incoming (post-makeup) sample is pushed
+     * into a [LIM_LOOKAHEAD]-sample delay line, and the gain is driven by the *newest*
+     * sample while it is applied to the *oldest* (delayed) one. When a sample louder
+     * than [LIM_CEILING] enters, the gain drops instantly to exactly the amount needed
+     * to tame it and is *held* for the whole look-ahead window (plus a small margin),
+     * so by the time that loud sample reaches the output the gain is already pulled
+     * down — the delayed peak can never exceed the ceiling. Between peaks the gain
+     * releases smoothly back toward unity, so quiet passages keep the full makeup.
+     * A final hard clamp is a cheap belt-and-braces guard (it should never engage).
+     */
+    private fun masterLimit(inL: Float, inR: Float) {
+        val dL = limDelayL[limPos]
+        val dR = limDelayR[limPos]
+        limDelayL[limPos] = inL
+        limDelayR[limPos] = inR
+        limPos++; if (limPos >= LIM_LOOKAHEAD) limPos = 0
+
+        val peak = max(abs(inL), abs(inR))
+        val target = if (peak > LIM_CEILING) LIM_CEILING / peak else 1f
+        when {
+            target < limGain -> { limGain = target; limHold = LIM_LOOKAHEAD + LIM_HOLD_EXTRA }
+            limHold > 0 -> limHold--
+            else -> limGain += (1f - limGain) * LIM_RELEASE
+        }
+
+        var oL = dL * limGain
+        var oR = dR * limGain
+        if (oL > LIM_CEILING) oL = LIM_CEILING else if (oL < -LIM_CEILING) oL = -LIM_CEILING
+        if (oR > LIM_CEILING) oR = LIM_CEILING else if (oR < -LIM_CEILING) oR = -LIM_CEILING
+        limOutL = if (oL.isFinite()) oL else 0f
+        limOutR = if (oR.isFinite()) oR else 0f
+    }
+
     /** Push slot parameters (+ tempo) into every live effect processor, but only when
      *  they actually changed — re-parsing unchanged params every block allocates and
      *  caused GC-pause dropouts when several effects were stacked. */
@@ -1225,10 +1284,20 @@ class AudioEngine(
         private const val LIVE_POLYPHONY = 8
         private const val AUDITION_POLYPHONY = 8
         private const val BUS_POLYPHONY = 8         // per-mixer-channel live audition
-        // Master make-up before the soft-clip limiter. Raised from 0.35 → 0.5 (~+3 dB)
-        // so playback sits closer to other media apps (it had noticeably more headroom
-        // than typical music playback); the soft-knee limiter below still catches peaks.
-        private const val MASTER_GAIN = 0.5f
+        // Master make-up gain into the look-ahead brickwall limiter ([masterLimit]).
+        // The old path was a flat 0.5× attenuation into a static soft-clip, which left
+        // ~9 dB of unused headroom: even the loudest single preset with its mix bus
+        // maxed peaked around -7 dBFS, well below other media playback. This pushes the
+        // signal UP (≈ +9.5 dB vs the old 0.5) so a single instrument reaches close to
+        // full scale, while the limiter guarantees the ceiling is never exceeded on
+        // busy passages — loudness without clipping. Tunable: raise for even louder /
+        // denser output, lower to trade loudness back for dynamics.
+        private const val MASTER_MAKEUP = 1.5f
+        // Brickwall limiter tuning (see [masterLimit]).
+        private const val LIM_CEILING = 0.97f        // hard output ceiling (~-0.26 dBFS)
+        private const val LIM_LOOKAHEAD = 64         // delay/look-ahead in samples (~1.3 ms @ 48 kHz)
+        private const val LIM_HOLD_EXTRA = 8         // extra hold beyond the window so the peak fully clears
+        private const val LIM_RELEASE = 0.00025f     // per-sample rise toward unity (~80 ms release @ 48 kHz)
         private const val LIVE_GAIN = 0.5f
         private const val SOFT_KNEE = 0.6f          // limiter is transparent below this, saturates above