⚡ export: screen-rect cursor plate, bounded unpremultiply, hwaccel decode

[devin-ai-integration]

📝 Description

Export profiling showed NVENC idle (~25% on RTX 3050) because every stage upstream is CPU-bound: software decode, a single-threaded zoompan, and a full-canvas RGBA overlay plate re-rasterized per frame in Rust and piped to FFmpeg's stdin as rawvideo (8–16 MB/frame). This PR does the cheap, no-restructure wins; the bigger reworks (pre-rendered overlay inputs, cursor-sprite sendcmd, chunked parallel export) stay as follow-ups.

Cursor plate shrinks to the fitted screen rect (cursor-only exports): the renderer still rasterizes into a canvas-sized pixmap (every draw is already clipped to video_screen), then crop_rgba_region extracts that rect and the graph composites it back via overlay=x={rx}:y={ry}. Per-frame rasterize→unpremultiply→pipe→composite work drops by the padding fraction (~10% at default padding, more for letterboxed/aspect-fit layouts). Dual-plane and items plates are unchanged — they draw in canvas coordinates.

unpremultiply_rgba_bounded restricts the premultiplied→straight-alpha pass to the bounding box of non-zero-alpha pixels. Pixels with alpha 0 are untouched by the conversion either way, so output is bit-identical while skipping ~90%+ of pixels on typical frames (a cursor sprite on a large canvas). Applied to both the cursor plane and the cached items plane.

-hwaccel auto on media inputs when a hardware encoder is selected: moves decode off the CPU where a backend exists (D3D11VA/DXVA2 on Windows, VideoToolbox on macOS, VAAPI on Linux), auto-downloads frames for the software filter graph, and falls back to software decode when no backend applies (verified on a GPU-less machine). Software encodes — including the existing hardware-failure retry, which re-enters with ExportEncoder::Software — still get a clean software command.

🎯 Type of Change

• ⚡ Performance optimization

---

🧪 Validation & Testing

• cargo fmt --check — clean

• cargo clippy -- -D warnings — clean

• cargo check — clean (stable-x86_64-pc-windows-gnu; this VM has no MSVC toolchain, so cargo test linking hits the GNU ld ordinal limit — left to CI)

• Standalone equivalence harness: unpremultiply_rgba_bounded produces byte-identical output to unpremultiply_rgba across sparse / dense / fully-transparent / single-edge-pixel / mixed-alpha inputs; crop_rgba_region verified pixel-exact

• -hwaccel auto verified to fall back to software decode cleanly on a GPU-less box (frame=60 ... rc=0)

• bun run typecheck — unchanged files are Rust-only; covered by CI

• bun run test — covered by CI

• cargo fmt --check / cargo clippy -D warnings / cargo check pass

• Tested on target operating systems:

  • Windows (Server 2022 VM; compile + unit-level verification — no GPU on this machine, so real NVENC utilization needs a run on actual hardware)
  • macOS
  • Linux

📋 Checklist

• Style follows existing conventions; cargo fmt clean
• Non-obvious decisions commented (rect-mirroring of ClipRect rounding, hwaccel gating rationale)
• No raw frames/secrets across Tauri boundaries
• No telemetry added

Expected impact: removes a chunk of the CPU work feeding NVENC (smaller plate + dirty-bbox unpremultiply + hw decode). GPU utilization should rise but likely stays well under 100% — zoompan remains single-threaded and the dual-plane path still ships full-canvas plates; raising GPU util substantially needs the chunked/sprite follow-ups.

Link to Devin session: https://app.devin.ai/sessions/52c7ba7f829f4aecbefa2f889efc8dc5
Open in Devin Desktop: https://app.devin.ai/desktop/session/52c7ba7f829f4aecbefa2f889efc8dc5?variant=devin
Requested by: @jeandedieuH

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring