初始提交：3 款纯音频游戏 + 共享引擎 + 完整文档与资产管线

- 共享引擎（app/src/main）：GestureEngine 6 手势识别、AudioEngine 三层音频（ExoPlayer BGM 交叉淡入 / 对白 ducking / SoundPool 低延迟）、Game/GameEngine 状态机接口、MainActivity 全屏封面 + TalkBack - 《声命线》storyline：末世互动广播剧，3 幕 8 抉择点 4 结局 - 《声东击西》reflex：听声辨位反应，12 波战斗 + BGM 三档 BPM 变速 - 《暗室》escape：3x3 网格密室逃脱，钥匙→铁柜→保险丝→电箱→大门谜题链 - 文档：setup（从零搭建）/ architecture（架构）/ engine-api（引擎 API）/ asset-pipeline（资产流水线）/ games/*（三款游戏设计文档） - 资产管线：Qwen3-TTS 对白 + 程序合成音效/BGM + Mureka BGM 的 gen 脚本与台账，全部 ogg 资产入库可复现 - env.d 环境脚本：android/tts 入库；sfx/mureka 含密钥不入库，仅提交 .example.sh 模板
2026-08-29 13:00:34 +08:00
commit f0f0a9b50e
446 changed files with 2627836 additions and 0 deletions
@@ -0,0 +1,107 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from tests.utils.audio import assert_audio_valid, sine_wave
+
+DURATION_SEC = 2
+
+
+def test_encode_decode(autoencoder):
+    ae = autoencoder
+    channels = ae.autoencoder.io_channels
+
+    audio = sine_wave(
+        DURATION_SEC, ae.sample_rate, channels=channels, device=ae.device
+    )  # (C, T)
+
+    latents = ae.encode(audio, ae.sample_rate)
+    recon = ae.decode(latents)
+
+    assert_audio_valid(recon, DURATION_SEC, ae.sample_rate)
+
+    # Loose reconstruction check: cosine similarity between flattened signals
+    min_len = min(audio.shape[-1], recon.shape[-1])
+    orig_flat = audio[..., :min_len].flatten().float()
+    recon_flat = recon[0, ..., :min_len].flatten().float()
+    similarity = F.cosine_similarity(
+        orig_flat.unsqueeze(0), recon_flat.unsqueeze(0)
+    ).item()
+    assert similarity > 0.5, (
+        f"Reconstruction cosine similarity too low: {similarity:.3f}"
+    )
+
+
+def test_preprocess_for_encoder(autoencoder):
+    """preprocess_audio_for_encoder resamples and pads a single (C, T) clip."""
+    ae = autoencoder
+    channels = ae.autoencoder.io_channels
+
+    # Generate at half the model rate to exercise resampling
+    input_sr = ae.sample_rate // 2
+    audio = sine_wave(DURATION_SEC, input_sr, channels=channels)  # (C, T)
+
+    batch = ae.autoencoder.preprocess_audio_for_encoder(audio, in_sr=input_sr)
+
+    assert batch.ndim == 3, f"Expected (1, C, T), got {batch.shape}"
+    assert batch.shape[0] == 1
+    assert batch.shape[1] == channels
+    # Output length should be a multiple of the downsampling ratio
+    assert batch.shape[2] % ae.autoencoder.downsampling_ratio == 0, (
+        f"Padded length {batch.shape[2]} is not a multiple of "
+        f"downsampling_ratio {ae.autoencoder.downsampling_ratio}"
+    )
+
+
+def test_preprocess_list_for_encoder(autoencoder):
+    """preprocess_audio_list_for_encoder handles clips of different lengths."""
+    ae = autoencoder
+    channels = ae.autoencoder.io_channels
+
+    clip_a = sine_wave(1, ae.sample_rate, channels=channels)  # 1 s
+    clip_b = sine_wave(DURATION_SEC, ae.sample_rate, channels=channels)  # 2 s
+
+    batch = ae.autoencoder.preprocess_audio_list_for_encoder(
+        [clip_a, clip_b], in_sr_list=ae.sample_rate
+    )
+
+    assert batch.ndim == 3, f"Expected (2, C, T), got {batch.shape}"
+    assert batch.shape[0] == 2
+    assert batch.shape[1] == channels
+    assert batch.shape[2] % ae.autoencoder.downsampling_ratio == 0
+
+
+def test_chunked_encode_decode(autoencoder):
+    """encode / decode with chunked=True produce valid audio."""
+    ae = autoencoder
+    channels = ae.autoencoder.io_channels
+
+    audio = sine_wave(
+        DURATION_SEC, ae.sample_rate, channels=channels, device=ae.device
+    )  # (C, T)
+
+    latents = ae.encode(audio, ae.sample_rate, chunked=True, chunk_size=128, overlap=32)
+    recon = ae.decode(latents, chunked=True, chunk_size=128, overlap=32)
+
+    assert_audio_valid(recon, DURATION_SEC, ae.sample_rate)
+
+
+def test_latent_save_load(autoencoder, tmp_path):
+    """Latents saved as .npy and reloaded decode to valid audio."""
+    ae = autoencoder
+    channels = ae.autoencoder.io_channels
+
+    audio = sine_wave(
+        DURATION_SEC, ae.sample_rate, channels=channels, device=ae.device
+    )  # (C, T)
+    latents = ae.encode(audio, ae.sample_rate)
+
+    # Save and reload via numpy
+    latent_path = tmp_path / "latents.npy"
+    np.save(latent_path, latents[0].cpu().numpy())
+
+    latent_tensor = torch.from_numpy(np.load(latent_path)).unsqueeze(0).to(ae.device)
+    assert latent_tensor.shape == latents.shape
+
+    recon = ae.decode(latent_tensor)
+    assert_audio_valid(recon, DURATION_SEC, ae.sample_rate)