diff --git a/main.lua b/main.lua
index 03b7abda..b9d3b4a9 100644
--- a/main.lua
+++ b/main.lua
@@ -521,9 +521,7 @@ function bootGame(version, cartId, opts)
   local loadOpts = { arena = arena, cartId = cartId, onExit = opts.onExit }
   if GameVersion.generation() == 3 then
     Game = require("src.core.Game3").new()
-    if arena then
-      Game.returnToLauncher = function(o) pendingLauncherReturn = o or {} end
-    end
+    Game.returnToLauncher = function(o) pendingLauncherReturn = o or {} end
     Game:load(loadOpts)
   elseif GameVersion.generation() == 2 then
     Game = require("src.core.Game2").new()
diff --git a/src/core/Game3.lua b/src/core/Game3.lua
index c747afd1..796ae541 100644
--- a/src/core/Game3.lua
+++ b/src/core/Game3.lua
@@ -227,6 +227,7 @@ function Game3:_handleBootAction(action)
     return
   end
   if action.action == "exit" then
+    Audio.stopAll()
     if self.returnToLauncher then
       self.returnToLauncher()
     elseif self.onExit then
diff --git a/src/core/game3/audio.lua b/src/core/game3/audio.lua
index 1ae28c68..febd7c98 100644
--- a/src/core/game3/audio.lua
+++ b/src/core/game3/audio.lua
@@ -36,13 +36,9 @@ Audio._bgmQueuedAt = {}
 Audio._bgmBaseAt = 0
 Audio._bgmVolume = 1
 Audio._sfxVolume = 1
--- Baked SE sit under BGM: CGB voice scales are conservative for the
--- shared mix bus, but doors/UI play as separate sources and need to cut through.
--- SE bake is quieter than hardware/mGBA mix under battle BGM; 3.5 keeps
--- move hits readable without hard-clipping typical effectiveness/move SE.
-Audio._seBakeGain = 3.5
 Audio._duck = 1
 Audio._duckHold = 0
+Audio._seDuck = 1
 Audio._mono = false
 Audio._warned = {}
 
@@ -190,8 +186,42 @@ function Audio.applyOptions(session)
   -- volumes reserved for future option fields
 end
 
-local function bgm_gain()
-  return (Audio._bgmVolume or 1) * (Audio._duck or 1) * (Audio._helpActive and 0.5 or 1)
+local function bgm_gain(volume)
+  return (volume or Audio._bgmVolume or 1)
+    * math.min(Audio._duck or 1, Audio._seDuck or 1)
+    * (Audio._helpActive and 0.5 or 1)
+end
+
+local function apply_bgm_gain()
+  if not Audio._bgmSource then return end
+  local gain = bgm_gain()
+  local f = Audio._fadeOut
+  if f and (f.gen == nil or f.gen == Audio._bgmGen)
+    and (f.songId == nil or (Audio._currentSong and f.songId == Audio._currentSong.id)) then
+    gain = bgm_gain(f.start) * (1 - math.min(1, f.t / math.max(f.dur, 0.01)))
+  elseif Audio._fadeIn then
+    f = Audio._fadeIn
+    gain = gain * math.min(1, f.t / math.max(f.dur, 0.01))
+  end
+  Audio._bgmSource:setVolume(gain)
+end
+
+-- Baked effects cannot steal a voice from already queued music. Until the
+-- players share a live mixer, give one-shot effects headroom at playback time.
+-- Looping alarms do not hold the music down; cry ducking takes precedence.
+local function update_se_duck(startingSource)
+  local active = false
+  for _, src in ipairs(Audio._seSources) do
+    local meta = Audio._seMeta[src]
+    if meta and meta.duckBgm and (src == startingSource or src:isPlaying()) then
+      active = true
+      break
+    end
+  end
+  local target = active and 0.6 or 1
+  local old = Audio._seDuck or 1
+  Audio._seDuck = target
+  if target ~= old then apply_bgm_gain() end
 end
 
 -- Help lowers BGM without changing the user's volume or cry ducking state.
@@ -328,7 +358,7 @@ function Audio.fadeOutBgm(speed)
     Audio._fadeOut = {
       t = 0,
       dur = seconds,
-      start = bgm_gain(),
+      start = Audio._bgmVolume or 1,
       songId = songId,
       gen = Audio._bgmGen,
     }
@@ -443,7 +473,7 @@ function Audio.playSe(id, opts)
 
   local pan = Audio.normalizePan(opts.pan)
   local sd = Player.bakeSlot(slot, {
-    master = (Audio._sfxVolume or 1) * (opts.volume or 1) * (Audio._seBakeGain or 3.5),
+    master = (Audio._sfxVolume or 1) * (opts.volume or 1),
     mono = Audio._mono,
     maxSec = opts.maxSec or ((loop or id == SE.SE_EXP) and 2.5 or 2.0),
     stopOnGoto = loop and true or false,
@@ -455,10 +485,14 @@ function Audio.playSe(id, opts)
     if loop then
       pcall(function() src:setLooping(true) end)
     end
-    src:play()
     Audio._seSources[#Audio._seSources + 1] = src
     Audio._seByPlayer[mplay] = src
-    Audio._seMeta[src] = { id = id, player = mplay }
+    Audio._seMeta[src] = { id = id, player = mplay,
+      duckBgm = not loop and id ~= SE.SE_SELECT
+        and (Audio._sfxVolume or 1) * (opts.volume or 1) > 0 }
+    update_se_duck(src)
+    src:play()
+    update_se_duck()
     while #Audio._seSources > 8 do
       local idx = 1
       if Audio._seSources[idx] == Audio._fanfareSource then idx = 2 end
@@ -491,6 +525,7 @@ function Audio.normalizePan(pan)
 end
 
 function Audio._songHasGoto(slot)
+  if slot and slot.info and slot.info.hasGoto ~= nil then return slot.info.hasGoto end
   local seq = slot and slot.seq
   if not seq or not seq.tracks then return false end
   for _, tr in ipairs(seq.tracks) do
@@ -509,6 +544,7 @@ function Audio._forgetSeSource(src)
     Audio._seByPlayer[meta.player] = nil
   end
   Audio._seMeta[src] = nil
+  update_se_duck()
 end
 
 function Audio._stopSePlayer(mplay)
@@ -706,17 +742,16 @@ function Audio.playCry(species, mode)
       rate = Mix.waveRate(meta.freq),
     })
     if src then
+      -- pokefirered PlayCry_Normal lowers BGM before starting the cry.
+      Audio._duck = 85 / 256
+      Audio._duckHold = 2
+      apply_bgm_gain()
       src:play()
       Audio._crySource = src
       local dur = (meta.size or 4000) / Mix.waveRate(meta.freq)
-      Audio._cryUntil = (Audio._cryClock or 0) + math.max(16, dur * 60)
+      Audio._cryUntil = (Audio._cryClock or 0) + dur * 60
     end
   end
-  -- Duck BGM to 85/256
-  Audio._duck = 85 / 256
-  Audio._duckHold = 2
-  if Audio._cmdCh then Audio._cmdCh:push({ cmd = "volume", volume = bgm_gain() }) end
-  if Audio._bgmSource then Audio._bgmSource:setVolume(bgm_gain()) end
   return true
 end
 
@@ -729,7 +764,7 @@ function Audio.isCryFinished()
   if Audio._cryUntil and (Audio._cryClock or 0) >= Audio._cryUntil then
     return true
   end
-  if Audio._crySource and Audio._crySource:isPlaying() then return false end
+  if Audio._crySource then return not Audio._crySource:isPlaying() end
   if not Audio._cryUntil then return true end
   return (Audio._cryClock or 0) >= Audio._cryUntil
 end
@@ -739,6 +774,8 @@ function Audio.stopCry()
   Audio._crySource = nil
   Audio._cryUntil = nil
   Audio._duck = 1
+  Audio._duckHold = 0
+  apply_bgm_gain()
 end
 
 function Audio.currentSong()
@@ -749,6 +786,7 @@ function Audio.stopAll()
   stop_bgm_source()
   Audio.stopSe()
   Audio.stopCry()
+  Audio._seDuck = 1
   Audio._currentSong = nil
   Audio._fanfareActive = false
 end
@@ -756,6 +794,7 @@ end
 function Audio.update(dt)
   dt = dt or 1 / 60
   Audio.tickCry(dt)
+  update_se_duck()
 
   Audio.pumpFanfares()
 
@@ -788,8 +827,7 @@ function Audio.update(dt)
     Audio._duckHold = Audio._duckHold - dt * 60
   elseif Audio._duck and Audio._duck < 1 and Audio.isCryFinished() then
     Audio._duck = 1
-    if Audio._cmdCh then Audio._cmdCh:push({ cmd = "volume", volume = bgm_gain() }) end
-    if Audio._bgmSource then Audio._bgmSource:setVolume(bgm_gain()) end
+    apply_bgm_gain()
   end
 
   -- Fade out/in
@@ -800,7 +838,7 @@ function Audio.update(dt)
     local stillSame = (f.gen == nil or f.gen == Audio._bgmGen)
       and (f.songId == nil or (Audio._currentSong and Audio._currentSong.id == f.songId))
     if stillSame and Audio._bgmSource then
-      local vol = f.start * (1 - u) * (Audio._duck or 1)
+      local vol = bgm_gain(f.start) * (1 - u)
       Audio._bgmSource:setVolume(vol)
     end
     if u >= 1 then
@@ -832,7 +870,7 @@ function Audio.update(dt)
     if okFree and type(free) == "number" and free > 0 then
       local n = Player.BUFFER_SAMPLES
       local sd = Player.renderBuffered(Audio._bgmLocal, n, {
-        master = bgm_gain(),
+        master = 1,
         sampleRate = Mix.SAMPLE_RATE,
       })
       if sd then
diff --git a/src/core/game3/m4a_mix.lua b/src/core/game3/m4a_mix.lua
index 886537ee..d9d95072 100644
--- a/src/core/game3/m4a_mix.lua
+++ b/src/core/game3/m4a_mix.lua
@@ -3,11 +3,12 @@
 
 local Mix = {}
 
--- FireRed m4aSoundMode: SOUND_MODE_FREQ_13379 → 224 PCM samples per VBlank.
-Mix.GBA_MIX_RATE = 13379
+-- FireRed SampleFreqSet: timer 0 advances every 1254 CPU cycles.
+Mix.GBA_CLOCK = 16777216
+Mix.GBA_FRAME_CYCLES = 280896
 Mix.GBA_SAMPLES_PER_VBLANK = 224
--- True GBA refresh: 13379/224 ≈ 59.7275 Hz (not 60).
-Mix.GBA_VBLANK_HZ = Mix.GBA_MIX_RATE / Mix.GBA_SAMPLES_PER_VBLANK
+Mix.GBA_VBLANK_HZ = Mix.GBA_CLOCK / Mix.GBA_FRAME_CYCLES
+Mix.GBA_MIX_RATE = Mix.GBA_VBLANK_HZ * Mix.GBA_SAMPLES_PER_VBLANK
 
 local function load_chip_synth()
   -- love.thread workers often cannot resolve package.path requires; mirror chip_worker.
@@ -110,10 +111,10 @@ local DS_FREQ = {
   3408917802, 3611622603, 3826380858, 4053909305,
 }
 
--- ChipSynth-style GB capacitor HPF (restores brightness vs DC-coupled bus).
+-- Remove DC from the unipolar GBA PSG without the DMG capacitor's bass loss.
 local GB_CLOCK = 4194304
 local function hpf_charge()
-  return 0.999958 ^ (GB_CLOCK / Mix.SAMPLE_RATE)
+  return math.exp(-2 * math.pi * 5 / Mix.SAMPLE_RATE)
 end
 
 local function midi_to_hz(key)
@@ -138,7 +139,7 @@ local function ds_scale_freq(key)
   return math.floor((DS_FREQ[s % 16] or 0) / (2 ^ math.floor(s / 16)))
 end
 
---- pret MidiKeyToFreq → Q10 sample rate (divide by 1024 for Hz).
+--- pret MidiKeyToFreq returns playback Hz; only WaveData.freq is Q10.
 function Mix.midiKeyToFreq(wavFreq, key, fine)
   wavFreq = tonumber(wavFreq) or 0
   key = math.floor(tonumber(key) or 60)
@@ -159,7 +160,7 @@ function Mix.midiKeyToFreq(wavFreq, key, fine)
 end
 
 --- pret MidiKeyToCgbFreq for pulse/wave → 11-bit-ish period register.
-function Mix.cgbPeriod(key, fine)
+function Mix.cgbPeriod(key, fine, fixed)
   key = tonumber(key) or 60
   fine = math.floor(tonumber(fine) or 0)
   if fine < 0 then fine = 0 end
@@ -178,19 +179,22 @@ function Mix.cgbPeriod(key, fine)
   local val1 = math.floor((CGB_FREQ[s1 % 16] or 0) / (2 ^ math.floor(s1 / 16)))
   local s2 = CGB_SCALE[key + 1] or s1
   local val2 = math.floor((CGB_FREQ[s2 % 16] or 0) / (2 ^ math.floor(s2 / 16)))
-  return val1 + math.floor((fine * (val2 - val1)) / 256) + 2048
+  local period = val1 + math.floor((fine * (val2 - val1)) / 256) + 2048
+  -- CgbSound quantizes FIX tones to the 65536 Hz PWM grid in FireRed's mode.
+  if fixed then period = math.floor((period + 1) / 2) * 2 % 2048 end
+  return period
 end
 
-function Mix.cgbPulseHz(key, fine)
-  local p = Mix.cgbPeriod(key, fine)
+function Mix.cgbPulseHz(key, fine, fixed)
+  local p = Mix.cgbPeriod(key, fine, fixed)
   local denom = 2048 - p
   if denom < 1 then denom = 1 end
   return 131072 / denom
 end
 
 -- Hardware wave clock is half of pulse → one octave lower for same period.
-function Mix.cgbWaveHz(key, fine)
-  return Mix.cgbPulseHz(key, fine) * 0.5
+function Mix.cgbWaveHz(key, fine, fixed)
+  return Mix.cgbPulseHz(key, fine, fixed) * 0.5
 end
 
 function Mix.periodToPulseHz(periodReg)
@@ -215,12 +219,8 @@ function Mix.cgbNoisePeriod(key)
   local nr43 = Mix.cgbNoiseNr43(key)
   local shift = math.floor(nr43 / 16) % 16
   local div = NOISE_DIV[nr43 % 8] or 8
-  local hz = 524288 / (div * (2 ^ shift))
-  if hz < 1 then hz = 1 end
-  local period = math.floor(Mix.SAMPLE_RATE / hz + 0.5)
-  if period < 1 then period = 1 end
-  if period > 4096 then period = 4096 end
-  return period
+  -- mGBA GB audio: divisor is already in 4 MHz clocks, not 524288 Hz units.
+  return Mix.SAMPLE_RATE * div * (2 ^ shift) / GB_CLOCK
 end
 
 --- Attach MP2K ADSR. CGB uses 0..15 sustain; DS uses 0..255.
@@ -243,114 +243,137 @@ function Mix.attachAdsr(voice, tone, isCgb, opts)
   return voice
 end
 
---- One sequencer-tick envelope step (pret SoundMain cadence ≈ per MPlayMain).
-function Mix.tickEnvelope(v)
-  if not v or not v.alive then return end
+local function finish_envelope(v, cgb)
   local a = v.adsr
-  if not a then
-    v.env = 1
-    return
-  end
-  local phase = v.envPhase or "attack"
-  if phase == "attack" then
-    if a.isCgb then
-      if (a.attack or 0) == 0 then
-        v.envVol = 15
-        v.envPhase = "decay"
-      else
-        v.envVol = math.min(15, (v.envVol or 0) + 1)
-        if v.envVol >= 15 then v.envPhase = "decay" end
-      end
-    else
-      if (a.attack or 0) >= 255 then
-        v.envVol = 255
-        v.envPhase = "decay"
-      else
-        v.envVol = (v.envVol or 0) + (a.attack or 0) + 1
-        if v.envVol >= 255 then
-          v.envVol = 255
-          v.envPhase = "decay"
-        end
-      end
-    end
-  elseif phase == "decay" then
-    if a.isCgb then
-      local sus = a.sustain or 0
-      if (a.decay or 0) == 0 then
-        v.envVol = sus
-        v.envPhase = "sustain"
-      else
-        v.envVol = (v.envVol or 15) - 1
-        if v.envVol <= sus then
-          v.envVol = sus
-          v.envPhase = "sustain"
-        end
-      end
-    else
-      local sus = a.sustain or 0
-      if (a.decay or 0) == 0 then
-        v.envVol = sus
-        v.envPhase = "sustain"
-      else
-        v.envVol = math.floor(((v.envVol or 255) * (a.decay or 0)) / 256)
-        if v.envVol <= sus then
-          v.envVol = sus
-          v.envPhase = "sustain"
-        end
-      end
-    end
-  elseif phase == "sustain" then
-    -- hold until gate → release
-  elseif phase == "release" then
-    if a.isCgb then
-      if (a.release or 0) == 0 then
-        v.envVol = 0
-        v.alive = false
-      else
-        v.envVol = (v.envVol or 0) - 1
-        if v.envVol <= 0 then
-          v.envVol = 0
-          v.alive = false
-        end
-      end
-    else
-      if (a.release or 0) == 0 then
-        v.envVol = 0
-        v.alive = false
-      else
-        v.envVol = math.floor(((v.envVol or 0) * (a.release or 0)) / 256)
-        local peVol = a.pseudoEchoVolume or 0
-        if peVol > 0 and (v.envVol or 0) <= peVol then
-          v.envVol = peVol
-          v.envPhase = "echo"
-        elseif (v.envVol or 0) <= 0 then
-          v.envVol = 0
-          v.alive = false
-        end
-      end
-    end
-  elseif phase == "echo" then
-    a.pseudoEchoLength = (a.pseudoEchoLength or 0) - 1
-    if a.pseudoEchoLength <= 0 then
-      v.envVol = 0
-      v.alive = false
-    end
-  end
-  local maxv = a.isCgb and 15 or 255
-  if maxv < 1 then maxv = 1 end
-  v.env = (v.envVol or 0) / maxv
-end
-
-function Mix.releaseVoice(v)
-  if not v then return end
-  if v.adsr then
-    v.envPhase = "release"
-    v.gateTicks = nil
+  local echo = a.pseudoEchoVolume or 0
+  if cgb then echo = math.floor(((v.envGoal or 0) * echo + 255) / 256) end
+  v.envVol = echo
+  if echo > 0 then
+    v.envPhase = "echo"
   else
     v.alive = false
   end
 end
 
+-- pokefirered CgbModVol: PSG has routing bits and a 4-bit envelope, not
+-- independent continuous left/right gains. Wave volume is quantized further.
+function Mix.cgbVolume(v)
+  local l = math.floor((v.volL or 0) * 256)
+  local r = math.floor((v.volR or 0) * 256)
+  v.routeL, v.routeR = true, true
+  if r >= l and math.floor(r / 2) >= l then
+    v.routeL = false
+  elseif l > r and math.floor(l / 2) >= r then
+    v.routeR = false
+  end
+  v.envGoal = math.min(15, math.floor((l + r) / 16))
+  v.sustainGoal = math.floor((v.envGoal * v.adsr.sustain + 15) / 16)
+end
+
+local function cgb_decay(v)
+  local a = v.adsr
+  v.envVol = v.envGoal
+  v.envPhase = "decay"
+  v.envCounter = a.decay
+  if a.decay == 0 then
+    v.envVol = v.sustainGoal
+    v.envPhase = "sustain"
+    v.envCounter = 7
+    if a.sustain == 0 then finish_envelope(v, true) end
+  end
+end
+
+local function tick_cgb_envelope(v, extraClock)
+  local a = v.adsr
+  Mix.cgbVolume(v)
+  if not v.envStarted then
+    v.envStarted = true
+    if v.released then v.alive = false; return end
+    v.envCounter = a.attack
+    if a.attack == 0 then cgb_decay(v) end
+  elseif v.envPhase == "echo" then
+    a.pseudoEchoLength = a.pseudoEchoLength - 1
+    if a.pseudoEchoLength <= 0 then v.alive = false end
+    return
+  elseif v.released and v.envPhase ~= "release" then
+    v.envPhase = "release"
+    v.envCounter = a.release
+    if a.release == 0 then finish_envelope(v, true) end
+  else
+    -- CgbSound counts down once per VBlank and twice every fifteenth frame.
+    if (v.envCounter or 0) == 0 then
+      if v.envPhase == "attack" then
+        v.envVol = v.envVol + 1
+        v.envCounter = a.attack
+        if v.envVol >= v.envGoal then cgb_decay(v) end
+      elseif v.envPhase == "decay" then
+        v.envVol = v.envVol - 1
+        v.envCounter = a.decay
+        if v.envVol <= v.sustainGoal then
+          v.envVol = v.sustainGoal
+          v.envPhase = "sustain"
+          v.envCounter = 7
+          if a.sustain == 0 then finish_envelope(v, true) end
+        end
+      elseif v.envPhase == "sustain" then
+        v.envVol = v.sustainGoal
+        v.envCounter = 7
+      elseif v.envPhase == "release" then
+        v.envVol = v.envVol - 1
+        v.envCounter = a.release
+        if v.envVol <= 0 then finish_envelope(v, true) end
+      end
+    end
+  end
+  if v.alive and v.envPhase ~= "echo" then
+    v.envCounter = math.max(0, (v.envCounter or 0) - 1)
+    if extraClock then tick_cgb_envelope(v, false) end
+  end
+end
+
+--- SoundMain envelope step, once per GBA frame regardless of song tempo.
+function Mix.tickEnvelope(v, extraCgbClock)
+  if not v or not v.alive then return end
+  local a = v.adsr
+  if not a then v.env = 1; return end
+  if a.isCgb then
+    tick_cgb_envelope(v, extraCgbClock)
+    v.env = (v.envVol or 0) / 15
+    return
+  end
+  if not v.envStarted then
+    v.envStarted = true
+    if v.released then v.alive = false; return end
+  end
+  local phase = v.envPhase
+  if phase == "echo" then
+    a.pseudoEchoLength = a.pseudoEchoLength - 1
+    if a.pseudoEchoLength <= 0 then v.alive = false end
+  elseif v.released then
+    v.envPhase = "release"
+    v.envVol = math.floor(v.envVol * a.release / 256)
+    if v.envVol <= a.pseudoEchoVolume then finish_envelope(v, false) end
+  elseif phase == "attack" then
+    v.envVol = math.min(255, v.envVol + a.attack)
+    if v.envVol == 255 then v.envPhase = "decay" end
+  elseif phase == "decay" then
+    v.envVol = math.floor(v.envVol * a.decay / 256)
+    if v.envVol <= a.sustain then
+      v.envVol = a.sustain
+      v.envPhase = "sustain"
+      if a.sustain == 0 then finish_envelope(v, false) end
+    end
+  end
+  v.env = v.envVol / 255
+end
+
+function Mix.releaseVoice(v)
+  if not v or v.released then return end
+  v.released = true
+  v.gateTicks = nil
+  if not v.adsr then v.alive = false end
+end
+
 function Mix.newDsVoice(pcm, meta, opts)
   opts = opts or {}
   local rate = opts.rate or Mix.waveRate(meta and meta.freq)
@@ -381,16 +404,18 @@ function Mix.waveRate(freqField)
 end
 
 -- Hardware: 4 exclusive CGB channels (pulse1, pulse2, wave, noise).
-Mix.MAX_DS_CHANNELS = 12
+Mix.MAX_DS_CHANNELS = 5
 
 function Mix.newCgbPulse(opts)
   opts = opts or {}
   local key = opts.key or 60
   local fine = opts.fine or 0
-  local periodReg = opts.periodReg or Mix.cgbPeriod(key, fine)
+  local fixed = opts.tone and math.floor((opts.tone.type or 0) / 8) % 2 == 1
+  local periodReg = opts.periodReg or Mix.cgbPeriod(key, fine, fixed)
   local v = {
     kind = "cgb_pulse",
     cgbChan = opts.cgbChan or 1, -- 1 or 2
+    cgbFixed = fixed,
     duty = opts.duty or 2,
     phase = 0,
     periodReg = periodReg,
@@ -422,12 +447,14 @@ function Mix.newCgbWave(opts)
   local wave = opts.wave or {}
   local key = opts.key or 60
   local fine = opts.fine or 0
+  local fixed = opts.tone and math.floor((opts.tone.type or 0) / 8) % 2 == 1
   local v = {
     kind = "cgb_wave",
     cgbChan = 3,
+    cgbFixed = fixed,
     wave = wave,
     phase = 0,
-    freq = opts.freq or Mix.cgbWaveHz(key, fine),
+    freq = opts.freq or Mix.cgbWaveHz(key, fine, fixed),
     volL = opts.volL or 0.35,
     volR = opts.volR or 0.35,
     env = opts.env or 1,
@@ -443,6 +470,7 @@ function Mix.newCgbNoise(opts)
     kind = "cgb_noise",
     cgbChan = 4,
     lfsr = 0x7FFF,
+    shortNoise = opts.tone and (tonumber(opts.tone.wavParam) or 0) % 2 == 1,
     clock = 0,
     period = opts.period or Mix.cgbNoisePeriod(opts.key or 60),
     volL = opts.volL or 0.3,
@@ -521,9 +549,10 @@ end
 
 local function render_voice(v, n, outL, outR)
   if not v or not v.alive then return end
-  local env = v.env
-  if env == nil then env = 1 end
-  if env <= 0 then return end
+  if v.lengthRemaining then
+    n = math.min(n, math.max(0, math.ceil(v.lengthRemaining)))
+    v.lengthRemaining = v.lengthRemaining - n
+  end
   if v.kind == "ds" then
     local pcm = v.pcm
     local size = v.size or #pcm
@@ -532,8 +561,10 @@ local function render_voice(v, n, outL, outR)
     local loopLen = size - loopStart
     local step = v.step or 1
     local pos = v.pos or 0
-    local volL = (v.volL or 0.5) * 0.45
-    local volR = (v.volR or 0.5) * 0.45
+    -- SoundMain: masterVolume = 12, envelope gain = ((12 + 1) * env) >> 4.
+    local e = math.floor(13 * (v.envVol or 255) / 16)
+    local volL = math.floor((v.volL or 0.5) * e) / 256
+    local volR = math.floor((v.volR or 0.5) * e) / 256
     for i = 1, n do
       if pos >= size then
         if loop and loopLen > 0 then
@@ -543,7 +574,7 @@ local function render_voice(v, n, outL, outR)
           break
         end
       end
-      local s = s8_lerp(pcm, pos, size, loop, loopStart) * env
+      local s = s8_lerp(pcm, pos, size, loop, loopStart)
       outL[i] = outL[i] + s * volL
       outR[i] = outR[i] + s * volR
       pos = pos + step
@@ -557,12 +588,12 @@ local function render_voice(v, n, outL, outR)
       if not v.alive then break end
       local inc = v.freq / Mix.SAMPLE_RATE
       local s = (v.phase < thresh) and hi or lo
-      -- Was 0.12 — left CGB SE (doors/select) ~3× under DS; pret CGB is louder.
-      s = s * env * 0.28
-      outL[i] = outL[i] + s * v.volL
-      outR[i] = outR[i] + s * v.volR
+      -- SOUNDCNT_L=0x77, PSG ratio=100%: one envelope unit is 16/512.
+      s = (s + 1) * (v.envVol or 15) / 64
+      if v.routeL ~= false then outL[i] = outL[i] + s end
+      if v.routeR ~= false then outR[i] = outR[i] + s end
       v.phase = v.phase + inc
-      if v.phase >= 1 then v.phase = v.phase - 1 end
+      v.phase = v.phase % 1
     end
   elseif v.kind == "cgb_wave" then
     local wave = v.wave
@@ -570,28 +601,82 @@ local function render_voice(v, n, outL, outR)
     for i = 1, n do
       local idx = math.floor(v.phase * 32) % 32
       local nibble = wave[idx + 1] or 8
-      local s = ((nibble / 7.5) - 1) * 0.25 * env
-      outL[i] = outL[i] + s * v.volL
-      outR[i] = outR[i] + s * v.volR
+      -- gCgb3Vol: mute, 25%, 50%, 75%, 100% (hardware truncates nibbles).
+      local level = v.envVol or 15
+      local sample = 0
+      if level >= 14 then sample = nibble
+      elseif level >= 10 then sample = math.floor(nibble * 3 / 4)
+      elseif level >= 6 then sample = math.floor(nibble / 2)
+      elseif level >= 2 then sample = math.floor(nibble / 4) end
+      local s = sample / 32
+      if v.routeL ~= false then outL[i] = outL[i] + s end
+      if v.routeR ~= false then outR[i] = outR[i] + s end
       v.phase = v.phase + inc
-      if v.phase >= 1 then v.phase = v.phase - 1 end
+      v.phase = v.phase % 1
     end
   elseif v.kind == "cgb_noise" then
     for i = 1, n do
       v.clock = v.clock + 1
-      if v.clock >= v.period then
-        v.clock = 0
+      while v.clock >= v.period do
+        v.clock = v.clock - v.period
         local lo = v.lfsr % 2
         local hi = math.floor(v.lfsr / 2) % 2
         local bit = (lo == hi) and 0 or 1
         v.lfsr = math.floor(v.lfsr / 2) + bit * 0x4000
+        if v.shortNoise then
+          v.lfsr = v.lfsr - (math.floor(v.lfsr / 64) % 2) * 64 + bit * 64
+        end
       end
-      local s = ((v.lfsr % 2) == 0) and 0.14 or -0.14
-      s = s * env
-      outL[i] = outL[i] + s * v.volL
-      outR[i] = outR[i] + s * v.volR
+      local s = ((v.lfsr % 2) == 0) and (v.envVol or 15) / 32 or 0
+      if v.routeL ~= false then outL[i] = outL[i] + s end
+      if v.routeR ~= false then outR[i] = outR[i] + s end
     end
   end
+  if v.lengthRemaining and v.lengthRemaining <= 0 then v.alive = false end
+end
+
+-- SoundMainRAM feeds the sum of both PCM sides, from six and seven frames
+-- ago, back into both sides at reverb/512. PSG never enters this buffer.
+-- The delay is resampled to the host clock; native s8 wrapping is not modeled.
+function Mix.newReverb(level, rate)
+  local frame = (rate or Mix.SAMPLE_RATE) / Mix.GBA_VBLANK_HZ
+  return {
+    level = math.max(0, math.min(127, tonumber(level) or 0)),
+    shortDelay = frame * 6,
+    longDelay = frame * 7,
+    size = math.ceil(frame * 7) + 1,
+    cursor = 0, left = {}, right = {}, energy = 0,
+  }
+end
+
+function Mix.reverbActive(state)
+  return state and state.level > 0 and state.energy > 1e-8
+end
+
+local function reverb_tap(state, delay)
+  local at = (state.cursor - delay) % state.size
+  local i = math.floor(at)
+  local frac = at - i
+  local j = (i + 1) % state.size
+  local a = (state.left[i] or 0) + (state.right[i] or 0)
+  local b = (state.left[j] or 0) + (state.right[j] or 0)
+  return a + (b - a) * frac
+end
+
+local function apply_reverb(state, left, right, n)
+  if not state or state.level <= 0 then return end
+  local gain = state.level / 512
+  for i = 1, n do
+    local echo = (reverb_tap(state, state.shortDelay)
+      + reverb_tap(state, state.longDelay)) * gain
+    local l, r = left[i] + echo, right[i] + echo
+    local at = state.cursor
+    local oldL, oldR = state.left[at] or 0, state.right[at] or 0
+    state.energy = math.max(0, state.energy + l * l + r * r - oldL * oldL - oldR * oldR)
+    state.left[at], state.right[at] = l, r
+    state.cursor = (at + 1) % state.size
+    left[i], right[i] = l, r
+  end
 end
 
 function Mix.render(voices, n, opts)
@@ -600,14 +685,18 @@ function Mix.render(voices, n, opts)
   local outL, outR = {}, {}
   for i = 1, n do outL[i] = 0; outR[i] = 0 end
   for _, v in ipairs(voices) do
-    render_voice(v, n, outL, outR)
+    if v.kind == "ds" then render_voice(v, n, outL, outR) end
+  end
+  apply_reverb(opts.reverbState, outL, outR, n)
+  for _, v in ipairs(voices) do
+    if v.kind ~= "ds" then render_voice(v, n, outL, outR) end
   end
   local alive = {}
   for _, v in ipairs(voices) do
     if v.alive then alive[#alive + 1] = v end
   end
 
-  -- Capacitor HPF (ChipSynth model) — restores highs vs a DC-coupled bus.
+  -- DC blocker; state belongs to the rendered player slot.
   local master = opts.master or 1
   local charge = hpf_charge()
   local capL = opts.hpfCapL
diff --git a/src/core/game3/m4a_player.lua b/src/core/game3/m4a_player.lua
index f1bff01f..42e96755 100644
--- a/src/core/game3/m4a_player.lua
+++ b/src/core/game3/m4a_player.lua
@@ -148,16 +148,11 @@ local function make_voice_from_tone(pack, tone, rawKey, absKey, volL, volR, fine
   if noteKey < 0 then noteKey = 0 end
   if noteKey > 178 then noteKey = 178 end
 
-  -- Rhythm pan override (if pan has bit 7 set)
-  if tone.isRhy and tone.pan and tone.pan >= 0x80 and tr then
-    local rpan = (tone.pan - 0xC0) * 2 -- -64..+62
-    local totalPan = ((tr.pan or 0x40) - 0x40) + rpan
-    if totalPan < -64 then totalPan = -64 elseif totalPan > 63 then totalPan = 63 end
-    local p = totalPan / 64
-    local vol = (tr.volume or 100) / 127 * ((tr.vel or 100) / 127)
-    volL = vol * (0.5 - p * 0.5)
-    volR = vol * (0.5 + p * 0.5)
+  local rhythmPan = 0
+  if tone.isRhy and tone.pan and tone.pan >= 0x80 then
+    rhythmPan = (tone.pan - 0xC0) * 2
   end
+  if tr then volL, volR = Seq.trackVolume(tr, tr.vel, rhythmPan) end
 
   local v
   if kind == 0 and tone.sampleId then
@@ -166,9 +161,10 @@ local function make_voice_from_tone(pack, tone, rawKey, absKey, volL, volR, fine
     if not pcm or not meta then return nil end
     local fixed = math.floor((tone.type or 0) / 8) % 2 == 1
     -- pret MidiKeyToFreq(wav, noteKey, fine) — returns playback Hz.
-    local rate = fixed and Mix.waveRate(meta.freq) or Mix.midiKeyToFreq(meta.freq, noteKey, fine)
-    if rate < 100 then rate = Mix.waveRate(meta.freq) end
-    local loop = (meta.loopStart or 0) > 0 and (meta.loopStart or 0) < (meta.size or 0)
+    local rate = fixed and Mix.GBA_MIX_RATE or Mix.midiKeyToFreq(meta.freq, noteKey, fine)
+    -- WaveData.flags is the high byte of cached status; loops may start at zero.
+    local loop = math.floor((meta.status or 0) / 16384) % 4 ~= 0
+      and (meta.loopStart or 0) < (meta.size or 0)
     v = Mix.newDsVoice(pcm, meta, {
       rate = rate,
       loop = loop,
@@ -216,6 +212,17 @@ local function make_voice_from_tone(pack, tone, rawKey, absKey, volL, volR, fine
   end
   if v then
     v.noteKey = baseKey
+    v.rhythmPan = rhythmPan
+    if v.adsr and tr then
+      v.adsr.pseudoEchoVolume = tr.pseudoEchoVolume or 0
+      v.adsr.pseudoEchoLength = tr.pseudoEchoLength or 0
+    end
+    if v.kind == "ds" then
+      v.pos = tr and tr.sampleStart or 0
+    elseif (tone.length or 0) > 0 then
+      local maxLength = kind == 3 and 256 or 64
+      v.lengthRemaining = (maxLength - tone.length % maxLength) * Mix.SAMPLE_RATE / 256
+    end
   end
   return v
 end
@@ -230,6 +237,12 @@ function Player.start(pack, cache, slot, songId, opts)
   slot.voices = {}
   slot.sampleOnly = false
   slot.done = false
+  slot.frameSamplesLeft = 0
+  slot.frameSampleCarry = 0
+  slot.hpfState = { l = 0, r = 0 }
+  local reverb = info and info.reverb or 0
+  if reverb >= 128 then slot.reverb = reverb % 128 end
+  slot.reverbState = Mix.newReverb(slot.reverb or 0)
 
   -- Prefer sequencer whenever track data exists. sampleOnly is only for
   -- explicit one-shots (cries) — SE like SE_SELECT are CGB sequences, not voice0 PCM.
@@ -271,6 +284,12 @@ function Player.start(pack, cache, slot, songId, opts)
         absKey = rawKey + keyM
       end
       local tone = resolve_tone(pack, vgId, voiceId, rawKey)
+      if tone and tr.toneOverrides then
+        local copy = {}
+        for k, value in pairs(tone) do copy[k] = value end
+        for k, value in pairs(tr.toneOverrides) do copy[k] = value end
+        tone = copy
+      end
       return make_voice_from_tone(pack, tone, rawKey, absKey, volL, volR, fine, tr)
     end,
   })
@@ -321,6 +340,7 @@ end
 
 function Player.renderSlot(slot, n, opts)
   opts = opts or {}
+  opts.reverbState = opts.reverbState or (slot and slot.reverbState)
   local voices = slot and slot.voices or {}
   if opts.raw then
     local outL, outR, alive = Mix.render(voices, n, opts)
@@ -339,20 +359,33 @@ function Player.renderBuffered(slot, n, opts)
   opts = opts or {}
   n = math.floor(tonumber(n) or Player.BUFFER_SAMPLES or 8192)
   if n < 1 then n = 1 end
-  local quantum = opts.quantum or Player.mixQuantum()
   local rate = opts.sampleRate or Mix.SAMPLE_RATE
   local master = opts.master or 1
   local L, R = {}, {}
   local produced = 0
   while produced < n do
-    local chunk = math.min(quantum, n - produced)
-    Player.updateSlot(slot, Mix.vblanksForSamples(chunk))
+    if (slot.frameSamplesLeft or 0) <= 0 then
+      Player.updateSlot(slot, 1)
+      local exact = rate / Mix.GBA_VBLANK_HZ + (slot.frameSampleCarry or 0)
+      slot.frameSamplesLeft = math.floor(exact)
+      slot.frameSampleCarry = exact - slot.frameSamplesLeft
+    end
+    local chunk = math.min(slot.frameSamplesLeft, n - produced)
+    slot.hpfState = slot.hpfState or { l = 0, r = 0 }
     local outL, outR, alive = Mix.render(slot.voices or {}, chunk, {
       raw = true,
       master = master,
       sampleRate = rate,
+      hpfCapL = slot.hpfState.l,
+      hpfCapR = slot.hpfState.r,
+      hpfState = slot.hpfState,
+      reverbState = slot.reverbState,
     })
-    if slot then slot.voices = alive or slot.voices end
+    if slot then
+      slot.voices = alive or slot.voices
+      if slot.seq then slot.seq.voices = slot.voices end
+      slot.frameSamplesLeft = slot.frameSamplesLeft - chunk
+    end
     for i = 1, chunk do
       L[#L + 1] = outL[i] or 0
       R[#R + 1] = outR[i] or 0
@@ -388,7 +421,8 @@ end
 
 function Player.snapshotSlot(slot, at)
   if not (slot and slot.seq) then return nil end
-  return { at = at, done = slot.done, seq = Seq.snapshot(slot.seq) }
+  return { at = at, done = slot.done, seq = Seq.snapshot(slot.seq),
+    frameSamplesLeft = slot.frameSamplesLeft, frameSampleCarry = slot.frameSampleCarry }
 end
 
 -- pokefirered/src/m4a.c:668
@@ -406,11 +440,13 @@ function Player.stopAt(slot, snaps, at, abs)
     slot.seq.voices = {}
     slot.voices = {}
     slot.done = snap.done
+    slot.frameSamplesLeft = snap.frameSamplesLeft or 0
+    slot.frameSampleCarry = snap.frameSampleCarry or 0
     local left = at - snap.at
     local q = Player.mixQuantum()
     while left > 0 do
       local n = math.min(q, left)
-      Player.updateSlot(slot, Mix.vblanksForSamples(n))
+      Player.renderBuffered(slot, n, { raw = true })
       left = left - n
     end
     for i = #snaps, 1, -1 do
@@ -421,6 +457,8 @@ function Player.stopAt(slot, snaps, at, abs)
   -- pokefirered/src/m4a_1.s:1469
   if slot.seq then slot.seq.voices = {} end
   slot.voices = {}
+  slot.hpfState = { l = 0, r = 0 }
+  slot.reverbState = Mix.newReverb(slot.reverb or 0)
   return abs
 end
 
@@ -431,21 +469,16 @@ function Player.bakeSong(pack, cache, songId, opts)
   local rate = opts.sampleRate or Mix.SAMPLE_RATE
   local quantum = Player.mixQuantum()
   local maxN = math.floor(rate * (opts.maxSec or 8))
-  local myL, myR = 0, 0
   local L, R = {}, {}
   local total, idle, since = 0, 0, 0
   while total < maxN do
     local n = math.min(quantum, maxN - total)
-    local capL, capR = Mix._hpfCapL, Mix._hpfCapR
-    Mix._hpfCapL, Mix._hpfCapR = myL, myR
     local outL, outR = Player.renderBuffered(slot, n, {
       raw = true,
       master = opts.master or 1,
       sampleRate = rate,
       quantum = quantum,
     })
-    myL, myR = Mix._hpfCapL, Mix._hpfCapR
-    Mix._hpfCapL, Mix._hpfCapR = capL, capR
     for i = 1, n do
       L[#L + 1] = outL[i] or 0
       R[#R + 1] = outR[i] or 0
@@ -460,7 +493,7 @@ function Player.bakeSong(pack, cache, songId, opts)
     for _, v in ipairs(slot.voices or {}) do
       if v.alive then any = true; break end
     end
-    if slot.done and not any then
+    if slot.done and not any and not Mix.reverbActive(slot.reverbState) then
       idle = idle + 1
       if idle >= 2 then break end
     else
@@ -493,8 +526,6 @@ function Player.bakeSlot(slot, opts)
   local L, R = {}, {}
   local total = 0
   local idle = 0
-  -- Fresh HPF state per bake so SE one-shots aren't coloured by BGM capacitors.
-  Mix._hpfCapL, Mix._hpfCapR = 0, 0
   local stopOnGoto = opts.stopOnGoto == true
   local sawGoto = false
   while total < maxN do
@@ -503,19 +534,17 @@ function Player.bakeSlot(slot, opts)
     if stopOnGoto and slot.seq and slot.seq.tracks then
       pcs = {}
       for i, tr in ipairs(slot.seq.tracks) do
-        pcs[i] = tr.pc or 0
+        pcs[i] = tr.loopCount or 0
       end
     end
-    Player.updateSlot(slot, Mix.vblanksForSamples(n))
+    local outL, outR = Player.renderBuffered(slot, n, { raw = true, sampleRate = rate })
     if pcs then
       for i, tr in ipairs(slot.seq.tracks) do
-        if (tr.pc or 0) < (pcs[i] or 0) then
+        if (tr.loopCount or 0) > (pcs[i] or 0) then
           sawGoto = true
         end
       end
     end
-    local outL, outR, alive = Mix.render(slot.voices or {}, n, { raw = true })
-    slot.voices = alive or slot.voices
     for i = 1, n do
       L[#L + 1] = outL[i] or 0
       R[#R + 1] = outR[i] or 0
@@ -528,7 +557,7 @@ function Player.bakeSlot(slot, opts)
     for _, v in ipairs(slot.voices or {}) do
       if v.alive then any = true; break end
     end
-    if slot.done and not any then
+    if slot.done and not any and not Mix.reverbActive(slot.reverbState) then
       idle = idle + 1
       if idle >= 2 then break end
     else
diff --git a/src/core/game3/m4a_seq.lua b/src/core/game3/m4a_seq.lua
index 3a494948..779b57b7 100644
--- a/src/core/game3/m4a_seq.lua
+++ b/src/core/game3/m4a_seq.lua
@@ -43,6 +43,7 @@ function Seq.parseSongBin(blob)
     reverb = u8(blob, 3),
     voicegroup = u32le(blob, 4),
     trackData = {},
+    trackEntries = {},
   }
   if tracks == 0 then return out end
 
@@ -50,12 +51,15 @@ function Seq.parseSongBin(blob)
   local packed = first > 0 and first < #blob and first < 0x01000000
 
   if packed then
+    local stride = (out.blocks or 0) >= 128 and 12 or 8
+    out.blocks = (out.blocks or 0) % 128
     for t = 0, tracks - 1 do
-      local base = 8 + t * 8
+      local base = 8 + t * stride
       local off = u32le(blob, base) or 0
       local len = u32le(blob, base + 4) or 0
       if off > 0 and len > 0 and off + len <= #blob then
         out.trackData[t + 1] = blob:sub(off + 1, off + len)
+        out.trackEntries[t + 1] = stride == 12 and (u32le(blob, base + 8) or 0) or 0
       else
         out.trackData[t + 1] = ""
       end
@@ -133,28 +137,33 @@ function Seq.trackPitch(tr)
   return math.floor(x / 256), (x % 256)
 end
 
-local function calc_track_vol(tr, vel)
+-- pokefirered TrkVolPitSet + ChnVolSetAsm, retaining the integer truncation
+-- and the separate rhythm pan multiplier. Gains are channel bytes / 256.
+function Seq.trackVolume(tr, vel, rhythmPan)
   vel = tonumber(vel) or tr.vel or 100
-  local vol = (tr.volume or 100) / 127 * (vel / 127)
-  if (tr.modT or 0) == 1 and (tr.mod or 0) > 0 then
-    local modVol = math.max(0, math.min(255, (tr.modM or 0) + 128)) / 128
-    vol = vol * modVol
+  local x = math.floor((tr.volume or 100) * 64 / 32)
+  if (tr.modT or 0) == 1 then
+    x = math.floor(x * ((tr.modM or 0) + 128) / 128)
   end
-  local pan = ((tr.pan or 0x40) - 0x40) / 64
-  if (tr.modT or 0) == 2 and (tr.mod or 0) > 0 then
-    pan = pan + (tr.modM or 0) / 128
-  end
-  if pan < -1 then pan = -1 elseif pan > 1 then pan = 1 end
-  local volL = vol * (0.5 - pan * 0.5)
-  local volR = vol * (0.5 + pan * 0.5)
-  return volL, volR
+  local pan = 2 * ((tr.pan or 0x40) - 0x40)
+  if (tr.modT or 0) == 2 then pan = pan + (tr.modM or 0) end
+  pan = math.max(-128, math.min(127, pan))
+  local ml = math.floor((127 - pan) * x / 256) % 256
+  local mr = math.floor((128 + pan) * x / 256) % 256
+  local rp = rhythmPan or 0
+  local l = math.min(255, math.floor((127 - rp) * vel * ml / 16384))
+  local r = math.min(255, math.floor((128 + rp) * vel * mr / 16384))
+  return l / 256, r / 256
 end
 
+local calc_track_vol = Seq.trackVolume
+
 function Seq.newPlayer(song, opts)
   opts = opts or {}
   local tracks = {}
   for i = 1, (song and song.tracks) or 0 do
     local tr = new_track(song.trackData[i])
+    tr.pc = song.trackEntries and song.trackEntries[i] or 0
     tr.index = i
     tracks[i] = tr
   end
@@ -163,6 +172,8 @@ function Seq.newPlayer(song, opts)
     tracks = tracks,
     tempo = 150, -- tempoD
     tempoC = 0,
+    frameC = 0,
+    c15 = 0,
     voices = {},
     voiceResolver = opts.voiceResolver,
     muted = false,
@@ -191,7 +202,7 @@ local function track_read32(tr)
   return b0 + b1 * 256 + b2 * 65536 + b3 * 16777216
 end
 
-local MAX_DS = 12
+local MAX_DS = Mix.MAX_DS_CHANNELS
 
 --- pret MPlayMain: after BEND/VOL/PAN/MOD, TrkVolPitSet + rewrite active channel freqs.
 local function refresh_track_voices(player, tr)
@@ -201,7 +212,7 @@ local function refresh_track_voices(player, tr)
     if v.track == tr and v.alive ~= false then
       if tr._volDirty then
         local nvel = v.noteVel or tr.vel or 127
-        local volL, volR = calc_track_vol(tr, nvel)
+        local volL, volR = calc_track_vol(tr, nvel, v.rhythmPan)
         v.volL = volL
         v.volR = volR
       end
@@ -212,17 +223,15 @@ local function refresh_track_voices(player, tr)
         if absKey > 178 then absKey = 178 end
         if v.kind == "ds" and v.wavFreq then
           local rate = Mix.midiKeyToFreq(v.wavFreq, absKey, fine)
-          if rate >= 100 then
-            v.step = rate / Mix.SAMPLE_RATE
-          end
+          v.step = rate / Mix.SAMPLE_RATE
         elseif v.kind == "cgb_pulse" then
-          local period = Mix.cgbPeriod(absKey, fine)
+          local period = Mix.cgbPeriod(absKey, fine, v.cgbFixed)
           v.periodReg = period
           if not v.sweepEnabled then
             v.freq = Mix.periodToPulseHz(period)
           end
         elseif v.kind == "cgb_wave" then
-          v.freq = Mix.cgbWaveHz(absKey, fine)
+          v.freq = Mix.cgbWaveHz(absKey, fine, v.cgbFixed)
         elseif v.kind == "cgb_noise" then
           v.period = Mix.cgbNoisePeriod(absKey)
         end
@@ -235,20 +244,19 @@ end
 
 local function start_note(player, tr, key, vel, gate)
   vel = vel or tr.vel or 100
-  local volL, volR = calc_track_vol(tr, vel)
   if not player.voiceResolver then return end
+  if (tr.lfoDelay or 0) > 0 then
+    tr.lfoDelayC = tr.lfoDelay
+    tr.modM = 0
+    tr._pitchDirty, tr._volDirty = true, true
+  end
+  local volL, volR = calc_track_vol(tr, vel)
   local rawKey = tonumber(key) or 60
   local keyM, fine = Seq.trackPitch(tr)
   local absKey = rawKey + keyM
   if absKey < 0 then absKey = 0 end
   if absKey > 178 then absKey = 178 end
 
-  -- Reset LFO delay on new note per pret ply_note
-  if (tr.lfoDelay or 0) > 0 then
-    tr.lfoDelayC = tr.lfoDelay
-    tr.modM = 0
-  end
-
   -- Resolver gets standard args with absKey appended; voice stores base note key for live BEND.
   local voice = player.voiceResolver(tr.voice, rawKey, vel, tr, volL, volR, fine, absKey)
   if voice then
@@ -263,6 +271,8 @@ local function start_note(player, tr, key, vel, gate)
       voice.noteKey = rawKey
     end
     voice.noteVel = vel
+    voice.midiKey = rawKey
+    voice.priority = math.min(255, (player.song.priority or 0) + (tr.priority or 0))
 
     -- CGB: 4 physical hardware channels (1..4).
     -- pret MP2K (m4a_1.s lines 1648-1668): exactly one voice per CGB channel.
@@ -277,9 +287,9 @@ local function start_note(player, tr, key, vel, gate)
         end
       end
       if active then
-        local oldPrio = (active.track and active.track.priority) or 0
-        local newPrio = tr.priority or 0
-        if newPrio > oldPrio then
+        local oldPrio = active.priority or 0
+        local newPrio = voice.priority
+        if active.released or newPrio > oldPrio then
           active.alive = false
         elseif newPrio < oldPrio then
           return
@@ -294,16 +304,30 @@ local function start_note(player, tr, key, vel, gate)
         end
       end
     elseif voice.kind == "ds" then
-      -- DirectSound: pret maxChans (FireRed typically ≤12). Steal oldest.
-      local ds = {}
+      -- ply_note prefers a released voice, then the lowest priority and latest
+      -- track. A lower-priority note cannot evict a higher-priority active voice.
+      local count, victim = 0, nil
       for _, v in ipairs(player.voices) do
         if v.alive ~= false and v.kind == "ds" then
-          ds[#ds + 1] = v
+          count = count + 1
+          local eligible = v.released or (v.priority or 0) < voice.priority
+            or ((v.priority or 0) == voice.priority
+              and ((v.track and v.track.index) or 0) >= tr.index)
+          if eligible then
+            local vp, bp = v.priority or 0, victim and victim.priority or 0
+            local vi = (v.track and v.track.index) or 0
+            local bi = victim and victim.track and victim.track.index or 0
+            if not victim or (v.released and not victim.released)
+              or ((not not v.released) == (not not victim.released)
+                and (vp < bp or (vp == bp and vi >= bi))) then
+              victim = v
+            end
+          end
         end
       end
-      while #ds >= MAX_DS do
-        local victim = table.remove(ds, 1)
-        if victim then victim.alive = false end
+      if count >= MAX_DS then
+        if not victim then return end
+        victim.alive = false
       end
     end
 
@@ -347,23 +371,40 @@ end
 local function exec_cmd(player, tr, cmd)
   if cmd == 0xB1 then
     tr.done = true
+    for _, v in ipairs(player.voices) do
+      if v.track == tr then Mix.releaseVoice(v) end
+    end
   elseif cmd == 0xB2 then
     local addr = track_read32(tr)
-    tr.pc = (addr < #tr.data) and addr or 0
+    if addr < #tr.data then
+      if addr < tr.pc then tr.loopCount = (tr.loopCount or 0) + 1 end
+      tr.pc = addr
+    else
+      exec_cmd(player, tr, 0xB1)
+    end
   elseif cmd == 0xB3 then
     local addr = track_read32(tr)
-    tr.callStack[#tr.callStack + 1] = tr.pc
-    if addr < #tr.data then tr.pc = addr end
+    if addr < #tr.data and #tr.callStack < 3 then
+      tr.callStack[#tr.callStack + 1] = tr.pc
+      tr.pc = addr
+    else
+      exec_cmd(player, tr, 0xB1)
+    end
   elseif cmd == 0xB4 then
     if #tr.callStack > 0 then tr.pc = table.remove(tr.callStack) end
   elseif cmd == 0xB5 then
-    track_read(tr)
+    local count = track_read(tr) or 0
     local addr = track_read32(tr)
-    tr.callStack[#tr.callStack + 1] = tr.pc
-    if addr < #tr.data then tr.pc = addr end
+    tr.repN = (tr.repN or 0) + 1
+    if count == 0 or tr.repN < count then
+      if addr < #tr.data then tr.pc = addr end
+    else
+      tr.repN = 0
+    end
+  elseif cmd == 0xBA then
+    tr.priority = track_read(tr) or 0
   elseif cmd == 0xBB then
     local t = track_read(tr) or 75
-    if t < 1 then t = 75 end
     player.tempo = t * 2
   elseif cmd == 0xBC then
     tr.keyShift = track_read(tr) or 0
@@ -371,6 +412,7 @@ local function exec_cmd(player, tr, cmd)
     tr._pitchDirty = true
   elseif cmd == 0xBD then
     tr.voice = track_read(tr) or 0
+    tr.toneOverrides = nil
   elseif cmd == 0xBE then
     tr.volume = track_read(tr) or 100
     tr._volDirty = true
@@ -410,40 +452,75 @@ local function exec_cmd(player, tr, cmd)
     tr.tune = (track_read(tr) or 0x40) - 0x40
     tr._pitchDirty = true
   elseif cmd == 0xCD then
-    -- Extended commands: CD <op> <args…>. Arity from pret gXcmdTable.
     local xop = track_read(tr) or 0
-    local xargs = ({
-      [0] = 0,  -- xxx
-      [1] = 4,  -- xwave (pointer)
-      [2] = 1,  -- xtype
-      [3] = 0,  -- unused
-      [4] = 1,  -- xatta
-      [5] = 1,  -- xdeca
-      [6] = 1,  -- xsust
-      [7] = 1,  -- xrele
-      [8] = 1,  -- xiecv
-      [9] = 1,  -- xiecl
-      [10] = 1, -- xleng
-      [11] = 1, -- xswee
-      [12] = 2, -- xwait (u16)
-      [13] = 4, -- xcmd_0D
-    })[xop] or 1
-    for _ = 1, xargs do track_read(tr) end
-  elseif cmd == 0xCE then -- endtie: release this track's notes only
-    for _, v in ipairs(player.voices) do
-      if v.track == tr then
+    local fields = { [2] = "type", [4] = "attack", [5] = "decay",
+      [6] = "sustain", [7] = "release", [10] = "length", [11] = "pan" }
+    if fields[xop] then
+      tr.toneOverrides = tr.toneOverrides or {}
+      tr.toneOverrides[fields[xop]] = track_read(tr) or 0
+    elseif xop == 8 then
+      tr.pseudoEchoVolume = track_read(tr) or 0
+    elseif xop == 9 then
+      tr.pseudoEchoLength = track_read(tr) or 0
+    elseif xop == 12 then
+      local low = track_read(tr) or 0
+      local high = track_read(tr) or 0
+      tr.wait = low + high * 256
+    elseif xop == 13 then
+      tr.sampleStart = track_read32(tr)
+    elseif xop == 1 then
+      -- Raw ROM pointer is not represented by the existing sample cache.
+      track_read32(tr)
+    else
+      exec_cmd(player, tr, 0xB1)
+    end
+  elseif cmd == 0xCE then
+    local key = tr.data:byte(tr.pc + 1)
+    if key and key < 0x80 then
+      tr.pc = tr.pc + 1
+      tr.key = key
+    else
+      key = tr.key
+    end
+    -- ply_endtie releases the newest matching MIDI key, not the whole track.
+    for i = #player.voices, 1, -1 do
+      local v = player.voices[i]
+      if v.track == tr and v.midiKey == key and v.alive and not v.released then
         Mix.releaseVoice(v)
+        break
       end
     end
   elseif cmd >= 0xCF then
     ply_note(player, tr, cmd)
   elseif cmd >= 0x80 and cmd <= 0xB0 then
     tr.wait = clock_at(cmd - 0x80)
-  elseif cmd == 0xBA or cmd == 0xC6 or cmd == 0xC7
-      or cmd == 0xC9 or cmd == 0xCA or cmd == 0xCB or cmd == 0xCC
-      or cmd == 0xB9 then
-    -- PRIO/… (1 arg); runningStatus already set if >= 0xBD
+  elseif cmd == 0xCC then -- PORT register offset and value
     track_read(tr)
+    track_read(tr)
+  elseif cmd == 0xB9 then
+    local op = track_read(tr) or 0
+    local addr = track_read(tr) or 0
+    local data = track_read(tr) or 0
+    player.mem = player.mem or {}
+    local value = player.mem[addr] or 0
+    local operand = op >= 3 and op <= 5 or op >= 12
+    operand = operand and (player.mem[data] or 0) or data
+    if op <= 5 then
+      local action = op % 3
+      if action == 0 then value = operand
+      elseif action == 1 then value = value + operand
+      else value = value - operand end
+      player.mem[addr] = value % 256
+    elseif op <= 17 then
+      local target = track_read32(tr)
+      local cond = (op - 6) % 6
+      local take = (cond == 0 and value == operand) or (cond == 1 and value ~= operand)
+        or (cond == 2 and value > operand) or (cond == 3 and value >= operand)
+        or (cond == 4 and value <= operand) or (cond == 5 and value < operand)
+      if take and target < #tr.data then tr.pc = target end
+    end
+  else
+    exec_cmd(player, tr, 0xB1)
   end
 end
 
@@ -490,15 +567,15 @@ local function tick_track(player, tr)
 end
 
 local function seq_tick(player)
-  -- Age gates / envelopes at the start of the vblank (then new notes are added after).
-  for _, v in ipairs(player.voices) do
-    if v.gateTicks then
-      v.gateTicks = v.gateTicks - 1
-      if v.gateTicks <= 0 then
-        Mix.releaseVoice(v)
+  -- MPlayMain ages each track's gates immediately before its commands.
+  for _, tr in ipairs(player.tracks) do
+    for _, v in ipairs(player.voices) do
+      if v.track == tr and v.gateTicks then
+        v.gateTicks = v.gateTicks - 1
+        if v.gateTicks <= 0 then Mix.releaseVoice(v) end
       end
     end
-    Mix.tickEnvelope(v)
+    tick_track(player, tr)
   end
 
   -- Tick LFO modulation per active track
@@ -530,39 +607,32 @@ local function seq_tick(player)
     end
   end
 
-  for _, tr in ipairs(player.tracks) do
-    tick_track(player, tr)
-  end
-
   for _, tr in ipairs(player.tracks) do
     if tr._pitchDirty or tr._volDirty then
       refresh_track_voices(player, tr)
     end
   end
 
-  -- New notes get one envelope step so attack=instant voices are audible this frame.
-  for _, v in ipairs(player.voices) do
-    if v.adsr and v.envPhase == "attack" and (v.envVol or 0) == 0 then
-      Mix.tickEnvelope(v)
-    end
-  end
 end
 
---- Advance by GBA VBlank units (one MPlayMain ≈ one vblank at 13379 Hz / 224 spv).
+--- Advance complete SoundMain frames. Tempo only governs MPlayMain ticks.
 function Seq.update(player, vblanks)
   vblanks = tonumber(vblanks) or 1
   if not player or player.muted then return player and player.voices or {} end
   if vblanks <= 0 then return player.voices end
-
-  local tempoI = player.tempo or 150
-  player.tempoC = (player.tempoC or 0) + tempoI * vblanks
-  local guard = 0
-  while player.tempoC >= 150 and guard < 64 do
-    guard = guard + 1
-    player.tempoC = player.tempoC - 150
-    seq_tick(player)
+  player.frameC = (player.frameC or 0) + vblanks
+  while player.frameC >= 1 do
+    player.frameC = player.frameC - 1
+    player.tempoC = (player.tempoC or 0) + (player.tempo or 150)
+    while player.tempoC >= 150 do
+      player.tempoC = player.tempoC - 150
+      seq_tick(player)
+    end
+    player.c15 = (player.c15 == 0) and 14 or player.c15 - 1
+    for _, v in ipairs(player.voices) do
+      Mix.tickEnvelope(v, player.c15 == 0)
+    end
   end
-
   local alive = {}
   for _, v in ipairs(player.voices) do
     if v.alive ~= false then alive[#alive + 1] = v end
@@ -581,13 +651,19 @@ function Seq.snapshot(player)
         local cs = {}
         for j, pc in ipairs(v) do cs[j] = pc end
         t.callStack = cs
+      elseif k == "toneOverrides" then
+        t.toneOverrides = {}
+        for field, value in pairs(v) do t.toneOverrides[field] = value end
       elseif type(v) ~= "table" then
         t[k] = v
       end
     end
     tracks[i] = t
   end
-  return { tracks = tracks, tempo = player.tempo, tempoC = player.tempoC }
+  local mem = {}
+  for k, v in pairs(player.mem or {}) do mem[k] = v end
+  return { tracks = tracks, tempo = player.tempo, tempoC = player.tempoC,
+    frameC = player.frameC, c15 = player.c15, mem = mem }
 end
 
 function Seq.restore(player, snap)
@@ -604,10 +680,19 @@ function Seq.restore(player, snap)
       local cs = {}
       for j, pc in ipairs(t.callStack or {}) do cs[j] = pc end
       tr.callStack = cs
+      tr.toneOverrides = nil
+      if t.toneOverrides then
+        tr.toneOverrides = {}
+        for k, v in pairs(t.toneOverrides) do tr.toneOverrides[k] = v end
+      end
     end
   end
   player.tempo = snap.tempo
   player.tempoC = snap.tempoC
+  player.frameC = snap.frameC or 0
+  player.c15 = snap.c15 or 0
+  player.mem = {}
+  for k, v in pairs(snap.mem or {}) do player.mem[k] = v end
 end
 
 function Seq.allDone(player)
diff --git a/src/core/game3/m4a_worker.lua b/src/core/game3/m4a_worker.lua
index 0bf5c1f0..6a66eb69 100644
--- a/src/core/game3/m4a_worker.lua
+++ b/src/core/game3/m4a_worker.lua
@@ -90,7 +90,8 @@ local function apply_cmd(msg)
     priority = tonumber(msg.id)
   elseif msg.cmd == "play" then
     if not pack then return end
-    bgm = { voices = {}, seq = nil, songId = msg.id, muted = false, volume = bgm.volume or 1, abs = 0, epoch = msg.epoch }
+    bgm = { voices = {}, seq = nil, songId = msg.id, muted = false,
+      volume = bgm.volume or 1, reverb = bgm.reverb, abs = 0, epoch = msg.epoch }
     snaps = {}
     Player.start(pack, cache, bgm, msg.id, { forceSeq = true })
   elseif msg.cmd == "stop" then
@@ -160,14 +161,14 @@ while running do
       if #snaps > SNAP_KEEP then table.remove(snaps, 1) end
     end
     local sd = Player.renderBuffered(bgm, BUFFER, {
-      master = bgm.volume or 1,
+      master = 1,
       sampleRate = sampleRate,
     })
     bgm.abs = at + BUFFER
     if sd then
       outCh:push({ gen = bgm.songId, epoch = bgm.epoch, at = at, n = BUFFER, data = sd, rate = sampleRate })
     end
-    if bgm.done and #(bgm.voices or {}) == 0 then
+    if bgm.done and #(bgm.voices or {}) == 0 and not Mix.reverbActive(bgm.reverbState) then
       outCh:push({
         gen = bgm.songId,
         epoch = bgm.epoch,
diff --git a/src/import/CacheContract.lua b/src/import/CacheContract.lua
index d3e1c103..1088de26 100644
--- a/src/import/CacheContract.lua
+++ b/src/import/CacheContract.lua
@@ -19,7 +19,8 @@ CacheContract.VERSION_FORMAT = {
   crystal = "rom-cache-v12-crystal4:",
   -- engine/overworld/map_sprites.asm:181
   yellow = "rom-cache-v11-yellow1:",
-  firered = "rom-cache-v7-firered:",
+  -- v8: M4A tracks retain reachable patterns and explicit entry offsets.
+  firered = "rom-cache-v8-firered:",
 }
 CacheContract.MARKER_PATH = "rom-cache.complete"
 
diff --git a/src/import/gba/extract_audio.lua b/src/import/gba/extract_audio.lua
index 1d90736f..32f52fcb 100644
--- a/src/import/gba/extract_audio.lua
+++ b/src/import/gba/extract_audio.lua
@@ -342,68 +342,84 @@ local function dump_voicegroup(data, vgOff, sampleMap, sampleParts, sampleCursor
   return id
 end
 
---- Rewrite GOTO/PATT absolute ROM pointers to track-relative offsets.
-local function relocate_track_pointers(body, romBase)
-  if type(body) ~= "string" or #body == 0 or not romBase then return body end
-  local b = { body:byte(1, #body) }
-  local i = 1
-  local function u32_at(idx)
-    return b[idx] + b[idx + 1] * 256 + b[idx + 2] * 65536 + b[idx + 3] * 16777216
-  end
-  local function set_u32(idx, v)
-    v = math.max(0, math.floor(v))
-    b[idx] = v % 256
-    b[idx + 1] = math.floor(v / 256) % 256
-    b[idx + 2] = math.floor(v / 65536) % 256
-    b[idx + 3] = math.floor(v / 16777216) % 256
-  end
-  local function reloc(idx)
-    if idx + 3 > #b then return end
-    local ptr = u32_at(idx)
-    local off = gba_off(ptr)
-    if off and off >= romBase and off < romBase + #b then
-      set_u32(idx, off - romBase)
+-- Walk reachable bytecode, including patterns beyond FINE or before the entry.
+-- Operands may contain command-valued bytes, so only instruction boundaries
+-- can identify FINE or a pointer. Note/running-status operands are all < 0x80.
+local function collect_track(data, entry)
+  local pending, visited, pointers = { { pc = entry, pattern = false } }, {}, {}
+  local first, last = entry, entry
+  local budget = 65536
+  local hasGoto = false
+  local xargs = { [0]=0, 4, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 4 }
+  while #pending > 0 do
+    local branch = table.remove(pending)
+    local pc, pattern = branch.pc, branch.pattern
+    while pc and pc >= 0 and pc < #data do
+      local state = pc * 2 + (pattern and 1 or 0)
+      if visited[state] then break end
+      budget = budget - 1
+      if budget < 0 then error("M4A track exceeds bytecode limit") end
+      visited[state] = true
+      local cmd = ru8(data, pc)
+      local length, pointer, stop = 1, nil, false
+      if cmd == 0xB2 or cmd == 0xB3 then
+        length, pointer = 5, pc + 1
+        stop = cmd == 0xB2
+        if stop then hasGoto = true end
+      elseif cmd == 0xB5 then
+        length, pointer = 6, pc + 2
+      elseif cmd == 0xB4 then
+        -- Inline patterns fall through PEND on their first, uncalled pass.
+        stop = pattern
+      elseif cmd == 0xB9 then
+        length = 4
+        local op = ru8(data, pc + 1) or 0
+        if op >= 6 and op <= 17 then length, pointer = 8, pc + 4 end
+      elseif cmd == 0xCD then
+        local op = ru8(data, pc + 1) or 0
+        length = 2 + (xargs[op] or 0)
+        stop = op == 0 or op == 3 or xargs[op] == nil
+      elseif cmd == 0xCC then -- PORT: register offset, value
+        length = 3
+      elseif cmd >= 0xBA and cmd <= 0xC5 or cmd == 0xC8 then
+        length = 2
+      elseif cmd >= 0xB1 and cmd <= 0xB8 or cmd == 0xC6 or cmd == 0xC7
+        or cmd >= 0xC9 and cmd <= 0xCB then
+        stop = true -- FINE, PEND, or an unused command mapped to ply_fine
+      end
+      if pc + length > #data then error("Truncated M4A instruction") end
+      first = math.min(first, pc)
+      last = math.max(last, pc + length)
+      if last - first > 1024 * 1024 then error("M4A track span exceeds limit") end
+      if pointer then
+        local target = gba_off(ru32(data, pointer))
+        if not target or target >= #data then error("Invalid M4A branch pointer") end
+        pointers[pointer] = target
+        pending[#pending + 1] = { pc = target, pattern = cmd == 0xB3 or pattern }
+      end
+      if stop then break end
+      pc = pc + length
     end
   end
-  while i <= #b do
-    local cmd = b[i]
-    if cmd == 0xB2 or cmd == 0xB3 then -- GOTO / PATT
-      reloc(i + 1)
-      i = i + 5
-    elseif cmd == 0xB5 then -- REPT count, ptr
-      if i + 5 <= #b then reloc(i + 2) end
-      i = i + 6
-    elseif cmd == 0xB1 or cmd == 0xB4 then
-      i = i + 1
-    elseif cmd == 0xCF then -- EOT
-      i = i + 1
-    elseif cmd == 0xCE then -- XCMD + 2 args
-      i = i + 3
-    elseif cmd >= 0xD0 then
-      i = i + 1
-    elseif cmd >= 0x80 and cmd <= 0xB0 then
-      i = i + 1
-    elseif cmd < 0x80 then
-      i = i + 2 -- key + vel
-    elseif cmd >= 0xB9 and cmd <= 0xCC then
-      -- most take 1 data byte (TEMPO/VOICE/VOL/PAN/BEND/BENDR/LFO/…)
-      i = i + 2
-    else
-      i = i + 1
-    end
+  local function le32(value)
+    return string.char(value % 256, math.floor(value / 256) % 256,
+      math.floor(value / 65536) % 256, math.floor(value / 16777216) % 256)
   end
-  local parts = {}
-  local CHUNK = 512
-  for i = 1, #b, CHUNK do
-    local j = math.min(i + CHUNK - 1, #b)
-    parts[#parts + 1] = string.char(unpack(b, i, j))
+  local ordered = {}
+  for pointer in pairs(pointers) do ordered[#ordered + 1] = pointer end
+  table.sort(ordered)
+  local parts, cursor = {}, first
+  for _, pointer in ipairs(ordered) do
+    parts[#parts + 1] = data:sub(cursor + 1, pointer)
+    parts[#parts + 1] = le32(pointers[pointer] - first)
+    cursor = pointer + 4
   end
-  return table.concat(parts)
+  parts[#parts + 1] = data:sub(cursor + 1, last)
+  return table.concat(parts), entry - first, hasGoto
 end
 
---- Packed song bin: header + (offset,length)×tracks + tight track bodies.
--- Track table uses file-relative offsets (always < 1MiB), never ROM pointers.
--- GOTO/PATT targets are relocated to offsets within each track body.
+--- Packed song bin: header + (offset,length,entry) per track. Bit 7 of blocks
+-- marks the extended table; branch offsets and entries are relative to each body.
 local function dump_song_tracks(data, headerOff, songId, cache, root)
   local tracks = ru8(data, headerOff) or 0
   if tracks == 0 or tracks > 16 then tracks = 0 end
@@ -417,52 +433,25 @@ local function dump_song_tracks(data, headerOff, songId, cache, root)
     romOffs[t] = gba_off(ru32(data, headerOff + 8 + t * 4))
   end
 
-  local sorted = {}
-  for t = 0, tracks - 1 do
-    if romOffs[t] then
-      sorted[#sorted + 1] = { t = t, off = romOffs[t] }
-    end
-  end
-  table.sort(sorted, function(a, b) return a.off < b.off end)
-
-  local lenByTrack = {}
-  for i, e in ipairs(sorted) do
-    local nextOff = sorted[i + 1] and sorted[i + 1].off
-    local len
-    if nextOff and nextOff > e.off then
-      len = nextOff - e.off
-    else
-      len = 0
-      for j = 0, 16383 do
-        if e.off + j >= #data then break end
-        if ru8(data, e.off + j) == 0xB1 then -- FINE
-          len = j + 1
-          break
-        end
-      end
-      if len == 0 then len = math.min(4096, #data - e.off) end
-    end
-    lenByTrack[e.t] = math.max(1, len)
-  end
-
-  local headerSize = 8 + tracks * 8
+  local headerSize = 8 + tracks * 12
   local cursor = headerSize
-  local trackMeta = {}
-  local bodyParts = {}
+  local trackMeta, bodyParts = {}, {}
+  local hasGoto = false
   for t = 0, tracks - 1 do
     local off = romOffs[t]
-    local len = (off and lenByTrack[t]) or 0
-    trackMeta[t] = { offset = cursor, length = len }
-    if off and len > 0 then
-      local body = data:sub(off + 1, off + len)
-      body = relocate_track_pointers(body, off)
-      bodyParts[#bodyParts + 1] = body
-      cursor = cursor + len
+    local body, entry = "", 0
+    if off then
+      local loop
+      body, entry, loop = collect_track(data, off)
+      hasGoto = hasGoto or loop
     end
+    trackMeta[t] = { offset = cursor, length = #body, entry = entry }
+    bodyParts[#bodyParts + 1] = body
+    cursor = cursor + #body
   end
 
   local hdr = {
-    string.char(tracks % 256, blocks % 256, priority % 256, reverb % 256),
+    string.char(tracks % 256, blocks % 128 + 128, priority % 256, reverb % 256),
     string.char(
       vgPtr % 256,
       math.floor(vgPtr / 256) % 256,
@@ -472,15 +461,16 @@ local function dump_song_tracks(data, headerOff, songId, cache, root)
   }
   for t = 0, tracks - 1 do
     local m = trackMeta[t] or { offset = 0, length = 0 }
-    local o, l = m.offset, m.length
+    local o, l, e = m.offset, m.length, m.entry or 0
     hdr[#hdr + 1] = string.char(
       o % 256, math.floor(o / 256) % 256, math.floor(o / 65536) % 256, math.floor(o / 16777216) % 256,
-      l % 256, math.floor(l / 256) % 256, math.floor(l / 65536) % 256, math.floor(l / 16777216) % 256
+      l % 256, math.floor(l / 256) % 256, math.floor(l / 65536) % 256, math.floor(l / 16777216) % 256,
+      e % 256, math.floor(e / 256) % 256, math.floor(e / 65536) % 256, math.floor(e / 16777216) % 256
     )
   end
   local blob = table.concat(hdr) .. table.concat(bodyParts)
   write(cache, string.format("%s/songs/%d.bin", root, songId), blob)
-  return #blob, tracks
+  return #blob, tracks, hasGoto
 end
 
 local function build_map_songs(rom, data)
@@ -576,7 +566,8 @@ function ExtractAudio.run(rom, cache, opts)
         info.priority = ru8(data, headerOff + 2)
         info.reverb = ru8(data, headerOff + 3)
         info.loop = (info.kind == "bgm")
-        local nbytes = dump_song_tracks(data, headerOff, id, cache, root)
+        local nbytes, _, hasGoto = dump_song_tracks(data, headerOff, id, cache, root)
+        info.hasGoto = hasGoto
         info.songBytes = nbytes
         local sid = song_primary_sample(data, headerOff, sampleMap, sampleParts, sampleCursor)
         info.sampleId = sid
diff --git a/src/import/gba/versions.lua b/src/import/gba/versions.lua
index 2a2bbb4b..0803f023 100644
--- a/src/import/gba/versions.lua
+++ b/src/import/gba/versions.lua
@@ -28,7 +28,7 @@ Versions.NATIVE_VERSION = 5
 Versions.OW_VERSION = 1
 Versions.ANIM_VERSION = 1
 -- Audio pack (M4A banks / DirectSound samples / cries).
-Versions.AUDIO_VERSION = 5
+Versions.AUDIO_VERSION = 6
 -- FireRed USA 1.0 (BPRE) — located by structural scan (entry0 ms=me=0, SE_SELECT ms=me=2).
 Versions.AUDIO = {
   song_table = 0x4A32CC,   -- gSongTable file offset