Skip to content

Record your screen, ship a demo. Free and open-source, GPU-accelerated, no watermarks, no subscriptions. Windows, macOS, Linux. Actively maintained.

docs(perf): record the shader and background-cache fixes on a Mac mini M1

Etienne Lescot27ed325360a0
diff --git a/technical-documentation/engineering/rendering-performance.md b/technical-documentation/engineering/rendering-performance.md
index de7178fe1..93c2387c4 100644
--- a/technical-documentation/engineering/rendering-performance.md
+++ b/technical-documentation/engineering/rendering-performance.md
@@ -376,6 +376,35 @@ In the editor (a dev build of `main` with rc.13's native directory), the preview
No register counts: Radeon GPU Analyzer was not run, so the VGPR explanation under fxc is inferred from the Linux measurement and from this outcome, not measured.
+### The same fixes on Metal: Mac mini M1, 2026-10-03
+
+The four fixes reached Metal (`ps_main` with and without `LAYER_MODELS`, the corner fast path, `bg_cache`, the trail scissor) without a measurement. Measured here on a Mac mini M1 (8 GiB, macOS 26.5), same bench and scenario, Parsec connected but idle (`parsecd` at 1.4 % of a core). "main" is `0624c148` packaged locally with `electron-builder --mac dir --arm64` and installed as `/Applications/Openscreen.app`; its `crates/`, `src/` and `electron/` are identical to 2.0.0-rc.13's.
+
+| | 2.0.0-rc.12 | main `0624c148` |
+|---|---:|---:|
+| benchmark median (3 runs) | 20.60 s | 19.38 s |
+| MAD | 0.02 s | 0.004 s |
+| local floor | 18.17 s | 18.19 s |
+| **× its local floor** | **1.134** | **1.066** |
+| foreign load per run | 54 to 70 % | 52 to 69 % |
+| CPU-seconds | 28.8 | 27.5 |
+
+Both in one session, back to back; the closing control re-measured the floor at 18.18 and 18.19 s, so the machine did not drift. The foreign load is the Claude desktop app and WindowServer, the same for both. The M1 had paid far less than the Radeon 610M for the uber-shader (1.11× in the published 2.0.0-rc.12 submission, against 1.04× for 1.11.0-rc.1), and the fixes take back most of it: −6 % wall clock. 1.11.0-rc.1 was not re-measured in this session: its run was discarded when a Parsec session went live halfway through (39.9 s exports, the hardware encoder shared with the remote screen), so the comparison with it stays the published 1.04×, from another day. The Apple GPU's register allocation was not read; there is no counterpart to `RADV_DEBUG=shaderstats` in this record yet.
+
+**Pixels are not byte-identical on Metal, unlike Linux, and the difference is rounding.** Each build is stable run to run (decoded pixels and SEI-stripped bitstream identical across three exports), but rc.12 and main differ: PSNR Y 93.5 dB on average, 66.9 dB at worst, on the bench output. VideoToolbox spreads any input change over the rest of its 12-frame GOP and moves its rate control, so the compositor was compared without it: both addons rebuilt with a lossless `ffv1` candidate added locally (`OPENSCREEN_EXPORT_ENCODER=ffv1`, measurement only, never committed), swapped into one app bundle so the addon was the only difference, five projects exported:
+
+| project | frames that differ / 3600 | largest luma difference |
+|---|---:|---|
+| bench project (wallpaper, radius, shadow, zooms, motion blur, trail, webcam) | 918 | 1/255, at most 3 px per frame, on the screen's anti-aliased corners |
+| laptop frame + 3D cursor + click impact + blurred wallpaper | 2 147 | 1/255 on 2 139 frames; 8 frames up to 15/255 on at most 30 px, on the laptop screen's silhouette |
+| phone frame (dark) + 3D Prism Glow cursor + animated gradient | 1 660 | 1/255, one frame at 8/255 on 5 px |
+| monitor frame + click impact + colour background | 1 988 | 1/255 on 1 971 frames; 17 frames up to 20/255 on at most 159 px, along the monitor's bottom bezel |
+| window frame + full background blur | 42 | 1/255 |
+
+No element is missing or displaced: the device shadows, the 3D cursor and its shadow, the click impact and the trail are where rc.12 drew them. The new corner fast path is exact in arithmetic (out of the corners `(q0 + e) − e` is `q0`), but Metal compiles `newLibraryWithSource` with fast math on, so a reshaped expression rounds differently; and the 3D models now come from their own library, whose ray-marched silhouettes flip a few edge pixels. The software encoder path writes luma only on macOS (`nv12_to_yuv420p` in `pipeline_macos.rs` is a stub, documented as never exercised), so this comparison covers the Y plane and part of the chroma, not all of it.
+
+The editor preview was driven over CDP on the same two builds (not computer-use: nobody was at the machine to grant it): the static frame differs from rc.12 by 1 or 2 pixels of 1/255 at a screen corner; changing the wallpaper (image, gradient, colour and back), the background blur (50 % and back) and the aspect ratio (1:1 and back) gives back the exact starting frame each time; a static background stays constant over a second of playback, and an *Aurora* background changes over the same second, so the cache keeps animated backgrounds out.
+
## The GIF export path — 2026-10-07
[#952](https://github.com/getopenscreen/openscreen/issues/952) measured GIF at ~3 fps against ~84 for MP4 on the reference laptop. Measured here with `export_gif` end to end on a generated clip: `testsrc2`, 6 s at 60 fps, exported at 864×480 and 15 fps (90 frames), dithered, release build, on a Ryzen 7 5800X (8 cores, 16 threads). The reference laptop was not measured.
@@ -875,7 +904,7 @@ the bench runs on the reference machine.
## Known gaps
-- **The Metal layer shader was split like the WGSL one without a measurement, and D3D11 only on an AMD iGPU.** On the Ryzen 5 7520U under Windows, the split, the static-background cache and the scissored trail together took the export from 4.78× to 1.71× its floor, level with 1.11.0-rc.1's 1.72×. The benchmark's scoring exports were byte-identical, and so were the model and frame variants; only the blurred-background variants differed, invisibly ([the Windows A/B](#the-same-fix-on-windows--2026-10-03)). Still owed: an Intel iGPU under Windows, fxc's register counts (Radeon GPU Analyzer, `ps_main` against `ps_main_models`), and an A/B on the M1, where the published figures are 1.04× for 1.11.0-rc.1 and 1.11× for 2.0.0-rc.12. [The Linux section](#the-linux-export-path--2026-10-02) has the method: register counts first, then the export.
+- **The split shaders were measured on one AMD iGPU under Windows and on the M1, and no register counts were read on either.** On the Ryzen 5 7520U under Windows, the split, the static-background cache and the scissored trail together took the export from 4.78× to 1.71× its floor, level with 1.11.0-rc.1's 1.72×. The benchmark's scoring exports were byte-identical, and so were the model and frame variants; only the blurred-background variants differed, invisibly ([the Windows A/B](#the-same-fix-on-windows--2026-10-03)). On the Mac mini M1 they took it from 1.134× to 1.066× its floor, pixels within 1/255 apart from model silhouettes ([the Metal A/B](#the-same-fixes-on-metal-mac-mini-m1-2026-10-03)). Still owed: an Intel iGPU under Windows, fxc's register counts (Radeon GPU Analyzer, `ps_main` against `ps_main_models`), and the Apple GPU's register allocation. [The Linux section](#the-linux-export-path--2026-10-02) has the method: register counts first, then the export.
- **macOS export startup can cost 4 s, and nobody has reproduced it on demand.** Measured repeatedly at 4208–4502 ms between the CLI's `started` event and the first composed frame — 18 % of a 60 s export, 71 % of a 5 s one — then gone, on the same shipped binary, hours later (481 ms). It is not the compositor (init is 2.4 ms, runtime MSL compilation included), not the `<video>` metadata probes (13 ms and 6 ms), not the CLI prologue (24 ms total), and not the renderer entry point (measured at −0.1 %). It correlates with memory pressure on an 8 GiB machine — `387M unused / 2613M compressor` while it reproduced, `564M unused / 1837M compressor` after — which would fit faulting ~1.8 MB of module chunks out of a 274 MB `app.asar` while the compressor thrashes: seconds of wall clock, no CPU in either process, cost independent of the media. Untested. Recreating the pressure deliberately and watching it return is what would settle it, and then whether asar size is the lever.
- **10-bit and HEVC decode on macOS are unmeasured.** The export's decode predicate is `codec_id == H264 && format == YUV420P`, so both keep VideoToolbox untested. HEVC is the case most likely to invert the result, since its software decoder is materially more expensive. 10-bit needs work beyond the predicate first: `mac_frames::CpuFrames` converts to 8-bit NV12, so routing 10-bit through the software path would silently truncate — the predicate is currently what prevents that.
- **The macOS preview's decode backend has never been measured.** `DecodeIntent` splits preview from export precisely so the preview could keep the old arbitration; the export won on throughput, but the preview scrubs, where seek latency after `avcodec_flush_buffers` may matter more, and it shares the machine with the editor UI. Changing it without measuring it would be the same mistake the export change corrects.
diff --git a/technical-documentation/testing/manual-e2e-checklist.md b/technical-documentation/testing/manual-e2e-checklist.md
index de6dfdeaf..252bfff96 100644
--- a/technical-documentation/testing/manual-e2e-checklist.md
+++ b/technical-documentation/testing/manual-e2e-checklist.md
@@ -855,3 +855,4 @@ On macOS 15.2+ it opens at launch, until it has been closed once, while one of i
| 2026-10-03 | installed `v2.0.0-rc.13` (NSIS from the release, sha512 matched `latest.yml`), installed over rc.12 for all users, About reports `2.0.0-rc.13` | Windows 11 Home 26200, 1920×1080 @ 125 %, AMD Radeon iGPU, built-in USB2.0 HD UVC WebCam | Partial — 1 defect | **Scoped to the app changes between rc.8, the last computer-use pass, and rc.13**; the export path itself is covered by the rc.13 row above. Real OS mouse and keyboard input throughout; computer-use screenshots came back uniformly grey again, so the app was observed through `PrintWindow` on its own windows and DOM reads over CDP. **Defect: the webcam shows its first frame when a clip is entered past the webcam's last frame from another clip.** Two clips of one take (screen video 50.98 s, webcam 51.24 s, audio and clip 51.31 s): scrubbing from clip 2 to 51.2 s in clip 1 shows the screen's last frame, as 227b9b84 intends, but the webcam's frame 0 (2 of 2; ground truth from ffmpeg: an empty chair from 18.7 s to the end). Scrubbing to the same point from inside clip 1, or to 51.0 s from clip 2, shows the right webcam frame. Cause: `open_and_seek_clip` and `seek_pair` in `crates/compositor/src/live.rs` seek the webcam with `seek_to` and fall back to `seek_to(0.0)`, where `present_frame` uses `seek_to_or_last`; unchanged on `main`. Paused scrub only, in a ~70 ms window at the end of each clip. **Passed:** #961: one take with the countdown overlay, Record mode, editor work and an export, then closing the editor left 0 processes. #966: after rebinding *Add zoom* to Q and *Add audio* to J, the empty-lane hints read "Appuyez sur Q pour ajouter un zoom" and "Appuyez sur J pour ajouter un audio, V pour enregistrer une voix off", and Q adds a zoom. Wordmark menu in French: every row 31 px (one line), menu 261 px wide, "2.0.0-rc.13" in one piece. Preview: playback from 0:12 to 0:45.8 across a trim (0:15.0 → 0:18.5) and a 1.5× region with no canvas stall over 150 ms, 32–50 picture changes a second, wall time matching the edit; stops at 0:51.3 on the screen's and the webcam's last frames. #960 (#964): A → B → A with the playhead at 0:06, the preview is pixel-identical to before (mean difference 0, against 77.9 for project B). Background image → gradient → colour followed in the preview. #968: a media file whose read the OS refused (`icacls` deny on the test file) — *Regenerate* shows the translated "Impossible de lire l'audio de ce média…" hint inline and in the "Échec de la transcription" toast, without the IPC wrapper; the main process logs `MediaUnreadableError … Permission denied`. MP4 export from the GUI, 1752×1080 30 fps: 2930 frames, video 97.633 s against audio 97.666 s, zoom, background, webcam and cursor present in extracted frames, −18.3 LUFS. **Also seen:** the green "Aucune parole détectée" chip from the earlier successful run stays next to the unreadable hint; each clip of the take still carries 0.32 s of audio past its video (what is left of #942), and the export ends each clip at its video (97.63 s, against 98.22 s from the clip lengths the timeline shows). **Not run:** the HUD with no camera (#967; a camera is attached), the macOS HUD fix (c2a8559e), tray, microphone and system-audio content, GIF, macOS and Linux. |
| 2026-10-03 | installed `v2.0.0-rc.13`, CI-built DMG `Openscreen-macOS-Apple-Silicon-2.0.0-rc.13.dmg` from the GitHub Release (tag `018d22ab`), installed over rc.12, `spctl`: accepted, Notarized Developer ID (M4LK7C6S84), About reports `2.0.0-rc.13` | macOS 26.5 (25F71), Mac mini M1, 3840×2160, no camera, no microphone | Partial, no defect | **Scoped to the app changes between rc.8 and rc.13, on macOS**: the Windows row above covers the same diff on Windows and leaves out #967, c2a8559e and macOS, which this row takes. Real OS mouse and keyboard input through computer-use (screenshots worked on this Mac); Apple's picker is owned by Control Centre and cannot be granted, so the operator clicked *Share Entire Screen* for each take. **c2a8559e (HUD kept out of takes after the editor round trip):** take 1 picked from the HUD's record button (`excludedWindowIds: [582]`), stop, editor, Record mode showed the source as *Choose what to record*, *Start recording* reopened Apple's picker instead of reusing the pick, take 2 recorded with `excludedWindowIds: [639]`; frames extracted from both files (28:26 and 13:04) show no HUD where it sat. **#967 (no camera on this Mac):** with `camEnabled: true` and no camera identity in `recording-settings.json`, launch rewrote it to `false` and the HUD showed the camera off; same with a stale identity (`camDeviceId` + `camDeviceName` of an unplugged camera); clicking the toggle keeps it off with the "Camera access is blocked" toast; Record mode reads *Camera Off*. **#961:** closing the editor after take 2 (countdown overlay, Record mode handover) left 0 processes; so did closing it after a long editor session (eight project opens through the native panel, transcriptions, Media mode, an export, the export dialog's save panel cancelled). **#966:** with *Add zoom* bound to X in `shortcuts.json`, the empty zoom lane reads "Press X to add zoom", and "Appuyez sur X pour ajouter un zoom" / "Drücke X, …" in French and German. **Wordmark menu (3bee2d25):** French and German, every row on one line, "2.0.0-rc.13" in one piece. **#960 (b834416c):** project A with captions → project B → A again: captions show at once at 0:06.4, 0:10.2 and 1:05.2 without touching *Show captions* (0:05.3 shows none, a 4.77 to 5.57 s silence in the saved words). **Preview:** playback from 0:12 to the end across four transcript silence cuts, a 2× and a 16× region: picture and captions follow, the readout advances with the edit, stops at 1:06.2; no stall seen, 8 `[pipeline] décodage` opens for one seek and four cuts (consistent with a preload per cut; not compared with rc.12). **227b9b84 / 83fbcecc:** a two-clip project whose first clip is a file with 8 s of video and 10 s of audio (no camera): seeking from clip 2 to 0:09.5 shows clip 1's last frame, the same as 0:07.7, no stderr. A video-only file plays with its picture (e73097c8). **#968:** project whose first asset is `chmod 000`: `MediaUnreadableError … Permission denied` in the main log, clip 2 still transcribed, the asset card shows "Transcription failed" and "Couldn't read this media's audio. Make sure OpenScreen can access the folder it's in, then regenerate." without the IPC wrapper; after `chmod 644`, *Regenerate* runs again (not persisted as no-audio). **Metal compositor (8caaa20d, dd5e914b, 0624c148, 2fb994eb):** laptop frame, 3D cursor, click impact, image background blurred 53 %, Aurora, captions; MP4 1080p60 export of 3374 frames, video and audio 56.23 s against 56.21 s computed (66.154 s minus 5.09 s of cuts minus the 2× and 16× gains), full decode clean; extracted frames show the laptop with its shadow, the blurred background, the 3D cursor with its shadow, rounded screen corners and captions. During playback the preview followed image → image, blur 52 → 1 %, image → colour and back. **Also seen:** the playhead keeps its time when another project is opened (0:41.0 from A shown in B and C). With a hand-written `camDeviceName` and a null `camDeviceId`, the HUD keeps the camera on and the file at `true` (`useCameraHudSync` compares the empty selection with the empty id and returns early); the app always writes both together, so this state is artificial. Out of scope: a 10 s extract with speech under loud video sound transcribes as no speech in Auto (detected `en`, p 0.35) and as "*Musique*" in French. **Not run:** webcam and microphone (none on this Mac), so the Windows row's webcam-past-the-end defect cannot be checked here; click impact not confirmed (the fixture's cursor sidecar has no clicks, and none was identified around the nine clicks of take 2 at preview scale); Aurora motion on a plain colour; the 4× preview lag of the rc.12 row (not re-measured: this pass played 2× and 16× regions without timing the lag); `Esc`; tray; GIF; A/B against rc.12; Gatekeeper (the DMG came through `gh`, no quarantine flag); Intel. |
| 2026-10-03 | installed `v2.0.0-rc.15` (build run 37124242835, tag `ccd8b449`, NSIS sha512 matched `latest.yml`), installed over rc.12 for all users, About reports `2.0.0-rc.15` (`win32 x64 · nsis`, Electron 41.2.1, Chromium 146.0.7680.188, Node 24.14.1). `v2.0.0` was promoted from this candidate during the pass and differs from it only by the version bump (`eff86dcb`), so the results hold for the stable code | Windows 11 Home 26200, 1920×1080 @ 125 %, AMD Ryzen 5 7520U / Radeon iGPU, built-in USB2.0 HD UVC WebCam (best mode 1280×720@30), Microphone Array (AMD) | Partial — 6 defects | **Whole-file pass on Windows.** Real OS mouse and keyboard input throughout. Computer-use screenshots came back grey again, so the app was observed through `PrintWindow` on its own windows and DOM reads over CDP, and every picture and sound claim below was measured on the file with ffprobe/ffmpeg. **Defects:** (1) #1004, HUD drag at 125 %: the HUD window grows about 1.5 px per drag step (1180×882 → 1232×935 over ~43 steps) and the bar jumps by half the growth on release; after a long drag down the bar sat partly under the taskbar (centre y 1042, work area ends at 1032). (2) #1005, *Edit clip*: Reset, Cancel and Apply sit inside the scrolling body. In a 1240×1000 window (794 css px tall) the card is clipped above Apply, and a click where Apply should be lands on the backdrop, closing the dialog and discarding the edit. It fits when the window is maximized. (3) #1006, the rail's "Choose a clip to edit" menu renders under the timeline toolbar: Clip 2 and Clip 3 are covered (`elementFromPoint` returns `_tlToolbar`) and clicks on them do nothing. (4) #1007, a clip duplicated with Ctrl+C / Ctrl+V does not carry its anchored zoom or Full Camera region: `duplicateClip` in `src/lib/ai-edition/document/timeline.ts` copies only `trimRanges`. (5) #1008, zoom repel across a clip boundary: a zoom dragged into a neighbour that ends at its clip's end jumps past it into the next clip, shrinking from 12.97 to 8.23 s and leaving a zero-length fragment on the first clip. Ctrl+Z restores it. (6) #1009, the wordmark menu does not close on a click on the bare top bar. The header is a `-webkit-app-region: drag` area, which never delivers the document `mousedown` that `AppMenu` listens for. It does close on the preview, on a top-bar button and on the wordmark. **Passed: HUD and capture.** One HUD, `[content-protection] OFF` logged; horizontal ↔ vertical layout; every tooltip, idle and recording, beside or above the bar and unclipped; language menu (15 locales); camera and microphone toggles in one click; device settings: three inputs, level meter, live camera preview, camera quality row (1080p persisted across a relaunch; hand-edited 720p and an absent key both read 4K). Hide bar, then the tray overflow icon brings the HUD back; Quit leaves 0 processes; a second launch exits 0. With no source the record tooltip reads "Choose a screen or window to record" and one click opens the picker, then starts. While recording: source button disabled, toggles `aria-disabled`, gear inert, tray tooltip "Recording: Tout l'écran". Pause freezes the timer and resume advances it; restart starts a new helper and deletes the first file; cancel deletes the file and opens no editor. Stop opens the editor with "Recording added to a new project". Screen H.264 High bt709 1920×1080; AAC 48 kHz; webcam 1280×720 ~30 fps at 8 Mbit/s, the camera's best mode, not upscaled, matching `webcamFormat`. Microphone-only, system-audio-only and all-off takes behave; window source 1270×668, non-black; an odd 1001×601 window gives an even 992×596 file; a context menu inside the window is recorded and cut at its edge. Without the flag the HUD and an open Notes window are absent from the take. Record mode: rows, live camera preview, microphone meter, device switch, source modal with badge, *Start recording* hands over to the HUD; auto-zoom on gives one merged zoom over three clicks, off gives none, and the setting persists; editable cursor off hides the auto-zoom row and the Cursor facet; *Hide desktop icons* shows a bare wallpaper and restores the icons after. **Passed: editor.** Rename, Saved dot, tabs, chat panel toggle, tooltips with key chips, dividers; play, pause, seek playing and paused, ±1 frame arrows, stop at end; ruler click and drag, navigator narrow and pan, labels without collision, playhead within 0.5 px; Ctrl+S toast with a stable top bar. Clips: media card drag and *Add to timeline*, reorder, Edit clip grips, 1:1 crop held through a corner drag, Free, Cancel, pencil, delete. Trim, zoom (levels, custom, out-of-range message, focus drag, 3D modes), speed (presets, custom, 16× cap message, playback at 16×), annotation (text, sizes, plates, palette, Typewriter, image, arrow, blur mosaic and smooth, oval, clamped, dragged into padding); Ctrl+D, Ctrl+Z, Ctrl+Shift+Z. Modifiers inside a trim fire on the parked frame. Copy and paste of zoom, annotation and trim with their toasts and properties; an empty copy is a no-op; deleting a clip removes its modifiers; Edit clip clamps them; a Full Camera region across a junction splits and merges as clips move. Imported music: fragments across a junction with `offsetMs`, −18 dB and 1 s fades, gain, *Reset audio*; voiceover lands at the playhead. Transcript: clip order, silences, monotonic word seeks, live cue, Backspace skip and restore, silence trim; captions in four styles and six anchors; French translation with a *Display* row, and deleting it leaves the words untouched; typing in the packaged gate does nothing; a double-click correction survives *Regenerate as English* (model reused); 101 languages, detected language on the card. Composition: colour, gradient and one-colour gradient, wallpaper, animations move during playback only (21–23 distinct frames in 3 s, None 1), blur, 9:16 with Whole and Follow cursor, Auto padding with even borders, shadows, four frame styles in both themes. Camera layout: four presets, mirror, square, size (default 40, max 60), roundness, positions; background Cutout, Blur with an intensity slider, Custom image and colour, and the custom-colour mask reaches a 720p export; webcam crop corner zoom (124 %) and frame move. Audio facet: +6.5 dB raises the export from −17.2 to −12.7 LUFS with the peak held at −1.3 dBFS; reset returns 0 dB. Cursor facet: hide, style, size (to 5.8 of 6) and 3D cursor all show in the preview. Depth of field appears once a zoom has a 3D preset and changes the preview. Wordmark menu: rows as listed, About row version, arrows wrap both ways; Keyboard Shortcuts as one dialog: Z → Q changes the lane hint, Q adds a zoom, Z does nothing, reset restores Z; Ctrl+O; AI settings as one dialog; light and dark themes, dialogs included; French UI with one-line menu rows (#969), back to English; About lists the versions and channel, and Copy puts the same block on the clipboard; the HUD panel's version row and Check for Updates ("Checking…", disabled, the same dialog as the menu, re-enabled after). AI chat (deepseek): "add one zoom at 2x from 5 s to 8 s" lands as 5–8 s at 2.20× with the reason given, is undone by one Ctrl+Z and redone. Closing the editor quits the app and writes `editor-window.json` (`maximized: false`); after a relaunch the last project's settings and the main project's two clips, crop, regions, music, voiceover and captions are as left, and a paste before any copy changes nothing. A schema-4 project from July opens, migrates to schema 8, and keeps its clip, seven zooms and annotation. **Export:** MP4 1080p30 of the two-clip project, 10288 frames, audio and video within one frame, −16.0 LUFS and −1.4 dBTP, annotation upright, caption plate at x = 108 for a 10 % inset, Full Camera where mapped, music ducked about 10 dB under speech with fades at the track's own edges, muted track absent, voiceover not stretched by a 2× region. GIF 25 fps: cancel removes the partial file and leaves the existing destination untouched; the full render (~127 s, #952) gives 434 frames, infinite loop, consistent in ffmpeg and GDI+. **Also seen:** a take with a ~4 s pause had audio 1.09 s past its video, against 0.02–0.17 s without a pause; three dedicated paused takes gave 0.40, 0.15 and 0.03 s. The export ends each clip at its last video frame (10288 frames against the 10320 the progress counted), so the timeline and the export disagree by that tail. The editor reopens 4 px taller than it closed at 125 %. **Not run:** the A/V offset across a pause measured directly (the flash-and-beep rig never painted on the captured desktop), the HUD with no camera (needs the camera disabled in Windows), tray right-click menus (Update Settings, Stop Recording, Save Diagnostics: Explorer is granted click-only), the Alt Help menu (an Alt tap showed nothing, inconclusive), Ctrl/Shift+wheel on the timeline, cursor auto-hide, smoothing, motion blur and click effects, annotation animations other than Typewriter, Depth of field in an export, chat rewind, history and Smart cuts, a software-encoder take, the crackle and hole scan, multiple displays, 4K and 1440p cameras, 150 % DPI, macOS and Linux. |
+| 2026-10-03 | `main` @ `0624c148` (sources identical to `v2.0.0-rc.13` in `crates/`, `src/`, `electron/`), packaged locally (`electron-builder --mac dir --arm64`, ad hoc signed), `compositor_view.node` SHA256 `7c1d38d664d29fa78615c386aa86e53456f1e662289ad53a2a128d588b99d560`; baseline the installed `v2.0.0-rc.12` DMG (Developer ID) | macOS 26.5 (25F71), Mac mini M1, 1920×1080 @ 2× | Partial, no defect | **Shader split, corner fast path, background cache and trail scissor (#982, #983, #984) on Metal. Not a computer-use pass:** `request_access` was refused with nobody at the machine, so the editor was driven over CDP (synthesised clicks on the inspector, preview pixels read from the canvas). `cargo test -p openscreen-compositor --lib --tests`: 407 passed, 0 failed, `every_shader_entry_point_compiles` (with the `LAYER_MODELS` library) and the macOS device-frame and 3D-cursor tests included; the `tests/` render binaries are Windows/Linux only and ran 0. **Export** (`openscreen export`, MP4 1080p60): bench project plus four variants (laptop + 3D cursor + click impact + blur, phone dark + Prism Glow + Aurora gradient, monitor + click impact + colour, window + full blur), frames extracted with ffmpeg: device frames with their shadows, 3D cursor with its shadow, click impact, trail, rounded corners and backgrounds all present. Same projects exported losslessly (`ffv1` candidate added locally, measurement only) with the rc.12 and main addons in one bundle: luma differs by 1/255 on anti-aliased edges, and by up to 20/255 on at most 159 px of a few frames along the 3D models' silhouettes (Metal fast math, models now in their own library); nothing missing. **Preview:** wallpaper image → gradient → colour → back, blur 50 % → 0 %, aspect 1:1 → Auto each return the exact starting frame; a static background holds still over 1 s of playback, *Aurora* animates; laptop and phone frames draw their shadows. **Perf** (screen-recorder-benchmark `commons-upload`): rc.12 1.134× floor, main 1.066×, figures in `rendering-performance.md`. **Not covered:** real recording, HUD, tray, computer-use input, the GUI export dialog, Escape, a 1.11.0-rc.1 run in the same session (discarded: a Parsec session went live mid-run), Windows (the D3D11 port is still unrun on a real GPU) and Linux. |