Add 24-bit integer and 32-bit floating-point PCM .wav input/output support

docs/ARCHITECTURE.md

@@ -342,7 +342,7 @@ EOF
     --embedding models/Qwen3-Embedding-0.6B-Q8_0.gguf \
     --dit models/acestep-v15-base-Q8_0.gguf \
     --vae models/vae-BF16.gguf \
-    --wav
+    --output wav
 ```
 
 Available track names for lego, extract, and complete: `vocals`, `backing_vocals`,
@@ -711,8 +711,14 @@ LoRA:
 
 Output:
   Default: MP3 at 128 kbps. input.json -> input0.mp3, input1.mp3, ...
+  --output <format>       Output audio file format (default: mp3)
+                            Supported values: mp3, wav, wav16, wav24, wav32
+                              mp3: MPEG-1 Audio Layer III encoded audio
+                              wav/wav16: 16-bit signed-integer WAVE audio
+                              wav24: 24-bit signed-integer WAVE audio
+                              wav32: 32-bit IEEE floating-point WAVE audio
+                                (wav32 disables normalization & peak clip)
   --mp3-bitrate <kbps>    MP3 bitrate (default: 128)
-  --wav                   Output WAV instead of MP3
 
 Memory control:
   --vae-chunk <N>         Latent frames per tile (default: 256)
@@ -923,6 +929,13 @@ Output:
   -o <path>               Output file (auto-named if omitted)
   --q8                    Quantize latent to int8 (~13 kbit/s)
   --q4                    Quantize latent to int4 (~6.8 kbit/s)
+  --wav-format <fmt>      WAV audio format (default: wav16)
+                            Requires use of --decode
+                            Supported values: wav, wav16, wav24, wav32
+                              wav/wav16: 16-bit signed-integer WAVE audio
+                              wav24: 24-bit signed-integer WAVE audio
+                              wav32: 32-bit IEEE floating-point WAVE audio
+                                (wav32 disables normalization & peak clip)
 
 Output naming: song.wav -> song.latent (f32) or song.nac8 (Q8) or song.nac4 (Q4)
                song.latent -> song.wav
@@ -979,16 +992,24 @@ from output extension).
 ```
 Usage: mp3-codec -i <input> -o <o> [options]
 
-  -i <path>   Input file (WAV or MP3)
-  -o <path>   Output file (WAV or MP3)
-  -b <kbps>   Bitrate for MP3 encoding (default: 128)
+  -i <path>          Input file (WAV or MP3)
+  -o <path>          Output file (WAV or MP3)
+  -b <kbps>          Bitrate for MP3 encoding (default: 128)
+  --wav-format <fmt> WAV audio format (default: wav16)
+                       Requires use of -o with a .wav extension
+                       Supported values: wav, wav16, wav24, wav32
+                         wav/wav16: 16-bit signed-integer WAVE audio
+                         wav24: 24-bit signed-integer WAVE audio
+                         wav32: 32-bit IEEE floating-point WAVE audio
+                           (wav32 disables normalization & peak clip)
 
 Mode is auto-detected from output extension.
 
 Examples:
   mp3-codec -i song.wav -o song.mp3
   mp3-codec -i song.wav -o song.mp3 -b 192
   mp3-codec -i song.mp3 -o song.wav
+  mp3-codec -i song.mp3 -o song.wav --wav-format wav32
 ```
 
 ## ace-understand reference

examples/lego.sh

@@ -21,7 +21,7 @@ set -eu
     --embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
     --dit ../models/acestep-v15-turbo-Q8_0.gguf \
     --vae ../models/vae-BF16.gguf \
-    --wav
+    --output wav
 
 # Phase 2: lego guitar on the generated track (base model required)
 ../build/ace-synth \
@@ -30,4 +30,4 @@ set -eu
     --embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
     --dit ../models/acestep-v15-base-Q8_0.gguf \
     --vae ../models/vae-BF16.gguf \
-    --wav
+    --output wav