Merge v0.5.1 + export fixes, GPU picker(nvidia/intel), NVENC, 4K presets

[progerWolf]

Merges v0.5.1 and adds work done on an NVIDIA Optimus laptop while cutting a 4K vertical video. It is a mix of general bug fixes, hardware-specific features, and packaging — split below so you can take, question or drop each part on its own.

Everything was verified against a running editor over CDP on both GPUs of this machine (NVIDIA GTX 1050 discrete, Intel UHD 630 integrated). npm run typecheck and npm run build are clean; e2e24, e2e28, e2e32, e2e35 and the node checks (proxy, envelope, ttstext, voicemap) pass, with one known exception noted at the end.

---

1. Bug fixes that affect everyone

These are not about my hardware. Each was reproduced before the change and measured after.

Audio: a clip with inPoint > 0 jumped to t=0 in the export

electron/ffmpeg.ts

ExportMuxer ended each segment's chain with an unbounded apad followed by atrim=0:<duration>. On an input-seeked segment (-ss inPoint -i file, i.e. any clip trimmed at the head) that trailing atrim threw the delayed audio back to t=0 and left its real slot silent. In a project cut into many pieces, most of the sound piled up at the start.

It is not a PTS problem: re-zeroing the timestamps (asetpts=PTS-STARTPTS, even asetpts=N/SR/TB) with the atrim still in place still collapsed the clip, while every chain without the trailing atrim placed it correctly.

Fix: apad + atrim → apad=whole_dur=<duration>. Each stream is padded to exactly the export length and then EOFs, so amix (duration=longest) still terminates — a two-input mix completes in 50 ms — and nothing trims after adelay.

Verified on a real project: clips at 25.667 s (inPoint 4.219) and 28.550 s (inPoint 3.033) land in place; the exported file's 7.5-second silent hole is gone.

Alpha: the matte was range-compressed twice in the packed intermediate

electron/ffmpeg.ts, electron/cacheKeys.ts

The packed colour-over-matte intermediate built its matte with alphaextract,format=yuv420p,scale=in_range=pc:out_range=tv. On current swscale (system ffmpeg 8.1 and a recent BtbN build alike) format=yuv420p already maps the gray alpha into limited range, so the explicit pc→tv after it compressed a second time: α 255 landed at Y=218 instead of 235. Every opaque area of an alpha source composited at ~92 % — a 255 red came out at 229.

Older swscale copied gray→yuv420p verbatim, which is why e2e32 passed where it was written.

Fix: put the range compression before format=. New swscale then sees a tv-tagged input and leaves it alone; old swscale still gets its single compression — 235/127/71 for α 255/129/64 on both ffmpeg builds here. The packed cache suffix moves :p → :p2 so existing intermediates re-pack.

Measured (e2e32): opaque red 229 → 250 (element path 251), PSNR 37.3 → 44.3 dB against the suite's own 40 dB gate. Identical numbers on NVIDIA and Intel — two very different decode paths — which is what shows the defect was in the packer, not the GPU.

Export intermediates went to /tmp, which is a tmpfs on many Linux setups

electron/runtime-env.ts, electron/main.ts, electron/preload.ts

The raw-video encode (preload) and the mux temp (main) used os.tmpdir(). Here /tmp is a 7.7 GB RAM-backed tmpfs; a large export overflows it mid-run and ffmpeg dies with EDQUOT/ENOSPC — surfacing as an export that stalls right after the start, or at ~67 % once the intermediate has grown. It cost me most of a day before I found the real error, because the UI just showed a frozen percentage.

runtime-env.ts now points KADR_TMPDIR at <userData>/tmp; both call sites use it with an os.tmpdir() fallback, and the renderer inherits the env the same way it already inherits KADR_FFMPEG.

Cancelling an export filed three failures in the session log

src/engine/exporter.ts, electron/preload.ts

Every cancelled export produced, on a run that ended exactly as asked:

Uncaught Error: write EPIPE
необработанный отказ промиса   (x2)

The two rejections are the frames still in flight: the export keeps two raw buffers in the air (slotPending) and awaits a slot only before reusing it, so killing the encoder rejects both with nothing left to await them. The finally block now claims them. The EPIPE is ffmpeg's stdin breaking under SIGKILL — the stream had no error listener, so Node escalated it to an uncaught exception, while the same failure was already delivered through the write callback and rawEncExit.

Verified live: export a clip, cancel mid-flight, count new log rows — 3 → 0, and e2e35 still passes all 16 checks, so a genuine context loss is still reported loudly.

This one matters beyond the noise: the panel added in v0.5.0 exists to surface real failures, and crying wolf on a routine action teaches people to ignore it.

An export could hang instead of failing

src/engine/demux.ts, electron/main.ts

Mp4FrameSource.open() had no timeout — a wedged decoder init hung there with neither an output nor an error, freezing the export on a frozen percentage. It gets a 10 s watchdog and falls back to element-seek, alongside an 8 s watchdog on the per-frame wait. Startup also sweeps orphaned kadr-export[-raw]-*.mp4 intermediates left by a crash (age-guarded, so a concurrent live export is never touched).

A keyframe without easing blacked out the whole frame, silently

src/engine/anim.ts, src/state/store.ts, src/engine/player.ts

evalAnim indexed ease[k0.easing] unguarded, so a keyframe written by a script or an imported project with a missing/unknown easing threw TypeError. Because comp.begin() clears the framebuffer before the layer loop, that one throw dropped every layer — a fully black frame, with the preview swallowing the error and the export aborting.

Three coordinated changes: evalAnim falls back to linear, sanitizeProject coerces unknown easings on load (healing existing project files), and drawFrame guards each clip so a bad one skips itself instead of taking the frame down.

---

2. Features

Take or leave — these are additions, not repairs.

• GPU picker (electron/gpu.ts, SettingsDialog) — choose the render GPU. On an Optimus laptop the discrete card needs PRIME offload set from process start, so picking NVIDIA re-execs with the offload env and --ozone-platform=x11; setting the vars late leaves the GPU process on the iGPU while GLX is forced to NVIDIA, which segfaults it. A pick is applied as a trial that self-heals to auto if the window never renders.
• NVENC export (shared/rawEncode.ts, ExportDialog) — optional hardware H.264, auto-falling back to x264. Availability is probed with a real 1-frame encode rather than an -encoders grep, because a build can list h264_nvenc and still fail to open it on an older driver (NVENC API 13.1 needs driver ≥ 610).
• 4K presets — vertical 2160×3840 and square 2160×2160.
• Project format picker — choose the canvas at New Project, and when the first video lands in an empty project ("match video" preselected).
• Rounded-rectangle mask — new shape with a rounded-box SDF in the compositor, a radius control, and a new squircle glyph (the icon set had none).
• Smaller things — re-importing a file overwritten under the same name refreshes the asset instead of reusing the stale one (dedupe by mtime); addTrack takes an optional insert index.

3. Packaging

scripts/pack.sh + electron-builder.yml build a self-contained archive (bundled ffmpeg/node/python/whisper, Remotion seed). Entirely additive; ignore it if you package differently.

One find worth flagging even if you drop the rest: four paths read at runtime through app.getAppPath() were missing from the packaged payload — electron/kadr-skill.md (missing since v0.3.0 introduced it), scripts/ttsqc_run.py, python/** and python/models/scorer.pkl. A packaged build starts fine and then fails the feature: the Claude panel logs skill sync failed: ENOENT, and voice-defect detection has no driver, package or model.

---

Decisions you may want to reverse

Fragments now render at the project fps instead of a forced ≥60. createFragment clamped to Math.max(60, fps). Remotion writes every frame as a full-res PNG to the OS temp dir, so on a 30 fps 4K project that doubled a pile that already overflowed a tmpfs — my renders died at frame 4877 of 5285. Since the timeline samples at the project fps, the extra frames are not visible in a 30 fps export.

This is a deliberate change to your choice, and it changed two assertions in e2e24 (a separate commit, f90f595). If you intended ≥60 for smoothness when exporting above the project fps, say so and I will make it conditional instead. Remotion's temp is also redirected off the tmpfs, with a pre-flight disk estimate that fails early with a clear message rather than at ~90 %.

Known failure, not fixed

e2e32's "both decode paths composite alpha identically" still fails: (250,0,0) vs (251,0,0) — 1–5 levels on saturated primaries, from rounding in the packed path's double matrix conversion against the element path's single one on newer swscale. Identical on both GPUs. The suite's own quality gate (PSNR ≥ 40) passes at 44.3 dB. I did not want to relax someone else's test on my own judgement — tell me which way you want it.

Not verified here

• Only Linux, only this machine (Bazzite, Wayland session, NVIDIA 580.173.02 + Intel UHD 630). No Windows or macOS.
• The GPU picker is Linux-only by construction (DRM render nodes, PRIME offload).
• The voice-over and defect features that arrived with v0.5.x were not exercised beyond confirming their runtime files now ship.

Happy to split this into separate PRs — the four fixes in §1 stand entirely on their own and do not depend on anything else here.