feat(export): linear audio mixdown for full-timeline export (refs Phase5/#112)

[Felix201209]

What

Extends the #112 full-timeline video export spine with a linear audio mixdown, turning the dead push_audio side-channel into a wired decode → mix → mux path. A timeline with audio now exports an mp4 with a real AAC track; a timeline without audio produces the same video-only file as before.

How

• opentake-media — new pure encode::mix module: lays each clip's mono f32 PCM at its frame-derived sample offset, applies a per-sample gain envelope, sums overlapping clips, and hard-limits to [-1.0, 1.0]. mono_f32_to_s16le produces the mux wire format. No ffmpeg/domain deps — fully unit-tested.
• opentake-media encoder: finish now mux's a supplied mixed buffer via a second ffmpeg pass (-c:v copy + AAC/LPCM, -shortest). On mux failure it restores the video-only file rather than losing the export. mux_args is pure and unit-tested; temp artifacts are cleaned up.
• export.rs orchestrator: decodes every audio/video clip's source window via the reused extract_pcm (mix rate, mono f32), builds the volume_at gain envelope, skips muted tracks and audio-less sources, mixes, and pushes to the encoder.
• export_integration.rs: ffmpeg+GPU-gated test builds a 440 Hz sine-audio fixture, exports, and asserts ffprobe reads back an AAC audio stream plus cleaned-up temps. The video-only test now also asserts no audio stream leaks in.

Scope

Linear mixdown skeleton only — no resampling curve, pan/stereo field, or dynamics (follow-ups). H.264/.mp4 path; H.265/ProRes still gated as before.

Verification

• cargo fmt --all clean
• cargo clippy -p opentake-media -p opentake-tauri --all-targets -- -D warnings clean
• cargo test -p opentake-media (24 encode/mix unit tests) green
• cargo test -p opentake-tauri export — 10 unit + 2 ffmpeg+GPU-gated integration tests green (audio test genuinely ran, not skipped)

Hard constraints honored: Tauri boundary returns Err(String); opentake-domain untouched (zero new deps); mixdown is pure logic in opentake-media; serde fields unchanged; unit tests under #[cfg(test)].

🤖 Generated with Claude Code

[appergb]

审核通过 ✅(自动审核流程 · 主控亲审,子 Agent 因 API 529 过载)

• #112 导出 spine 的线性音频混音扩展,真接通原 dead push_audio。
• mix.rs 纯混音正确:帧→样本偏移铺放、逐样本 volume_at 增益包络、叠加、硬限幅 [-1,1]、mono f32→s16le;12 单测覆盖空/单/偏移/重叠/溢出双轨/增益/长度校验/输出长度/编码与 clamp。
• encode finish() mux:第二趟 ffmpeg(-c:v copy + AAC/LPCM + -shortest);mux 失败回退视频-only 文件(不丢导出)、临时件 best-effort 清理;mux_args/sibling_temp 纯函数有测试。
• export.rs 编排:解码音/视频 clip 源窗(extract_pcm,mono f32 @48k)→增益包络→混音→push;跳过静音轨/无音源/NoTrack→静音;视频路径未改(-18 仅文档/import),无音频仍产出与 #112 相同视频-only 文件。
• 集成测试:加 440Hz 正弦 fixture 断言 AAC 轨 + 视频-only 断言不漏音轨 + 临时件清理。不碰 #91 区,CI 双绿。

follow-up(非阻塞):project_clip_audio ~60 行略超 50 规约;本切片仅线性混音(无重采样曲线/pan/立体声/动态),按设计后续。