Port main's live fixes into feat/ai-edition, and rewire the capabilities the V4 shell dropped

[EtienneLescot]

Rebuilt from the divergence point (68d3a685, 2026-07-15) rather than from the earlier rebase attempt, which is superseded and can be closed without merging (#154).

What the audit found

Comparing main against feat/ai-edition through the descriptions of the 15 PRs merged into main since the divergence, rather than a commit diff:

• main produced no new product feature in that window — only fixes and release tooling. The big ones (feat(editor): add Full Camera timeline effect #66 Full Camera, feat: add custom gradient editor to background settings #81 gradient editor, feat(wgc-capture): fall back to software H.264 encoder #73 software encoder, fix(preview, UI): resolve WebGL context loss and implement premium custom header menus #87 WebGL, fix(recording): parallelize screen/mic capture and fade in mic gain #58 parallel capture) landed before the divergence and are already here.
• But five of those fixes cover bugs still live in this branch.
• And separately: ancestry is not wiring. Several capabilities exist in this branch's code, fully implemented down to the compositor, but were never connected to the V4 shell.

Full write-up in docs/engineering/main-vs-ai-edition-audit.md and ai-edition-remediation-spec.md.

Shared fixes brought over (cherry-picks)

PR	Symptom
#122	Recording a window with odd client dimensions produced solid black video
#119	Under the software encoder, stopping a recording could never return
#123	Hardening for the above (unbounded per-frame wait)
#91	Linux/Wayland: capture produced no usable frames
#125	The HUD drifted away from the cursor while dragging

main.cpp was merged by hand to keep this branch's real-clock CFR webcam pacing while taking main's capture/submit split.

Bugs fixed here

The microphone was missing from exports. Native macOS recordings write system audio and mic as two separate AAC tracks; decode_clip_audio took one via av_find_best_stream — the first, silent whenever nothing plays through the system. This is issue #108, which PR #109 fixed in the browser exporter — a path this app no longer uses. Reimplemented in Rust: every audio stream decoded off a single demux pass, then mixed. 7 new tests.

Playback jumped back a clip at a reordered junction. A stale closure: the rAF loop is re-created only on asset swap, but called seekToVirtualTime, whose deps include clips. It picked the right next clip from fresh refs, then resolved it against the pre-reorder layout. Routed through the ref the seek effect already used.

A manual zoom drifted during its ease-in. The port fed lerp(0.5, f, strength) as the crop centre, which adds a scale factor the reference transform doesn't have — the focus point hangs away from centre mid-ramp, then snaps back. It reads as if a manual region were following the cursor. Fixed by inverting the crop mapping; a test sweeps the ease-in window and derives progress back out of the rendered scale.

Auto follow was jittery. cursorFollowUtils' adaptive exponential smoothing was never ported — only the sampling was. Ported, but applied once over the telemetry at load instead of per frame, so a rendered image never depends on the path taken to reach it (preview and export stay identical across seeks).

New text annotations were born filled with "New annotation", so the placeholder never showed.

Capabilities rewired to the V4 shell

Each already existed, with translations shipping in all 13 locales and no consumer:

• Speed above 3×. MAX_PLAYBACK_SPEED is 100; the inspector offered a preset select capped at 3. Restored the preset ladder plus a free field.
• Zoom focus mode. focusMode was read but never writable — "auto" only arrived from the suggestion pass.
• Global Auto-Focus toggle. The setting, its persistence and its labels all existed; the only control that wrote it lived in the unmounted Bottombar. While on, it overrides every region and the per-zoom control goes read-only with an explanation, so a zoom can no longer display "Manual" while the compositor follows the cursor. It never writes focusMode, so turning it off restores each zoom's own value.

Also removed the "above 16×" hint: it described the legacy editor's frame-stepped, silent preview, not this one.

Verification

tsc and biome clean. 1239 front-end tests across 101 files, 28 Rust tests — 20 of them new. Native helper and compositor addon both rebuilt and run.

Exercised in the running app: manual ease-in ramp, smoothed auto pan, the global toggle lock, multi-clip reorder, and webcam recording.

Not verified end to end: exporting a real multi-track macOS recording. The mixing logic is covered by unit tests, including the #108 case itself, but no such file was available to export.

Still open

Zoom rotation preset and non-text annotations remain unwired. preferSoftwareEncoder has no UI, and the D3D engine has no software fallback (D3D_DRIVER_TYPE_HARDWARE only) — both need a product call. MSIX packaging (#142) is ready on the superseded branch but pointless until this branch ships releases.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: bbc81492-0a58-4942-b717-4eb7038e06ab

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/ai-edition-remediation

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}