Skip to content

The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.

updates

Jareercf25b42f184b
diff --git a/ARCHITECTURE-TODO.md b/ARCHITECTURE-TODO.md
deleted file mode 100644
index f19a34f6..00000000
--- a/ARCHITECTURE-TODO.md
+++ /dev/null
@@ -1,41 +0,0 @@
-# Architecture: left for later
-
-What the 16 September architecture proposal ticked off with a note that
-part of it waits. The proposal itself is gone; `ARCHITECTURE.md` describes
-the tree as it is, and this is the list of what it does not yet do. Each
-item names where the seam is, so the next person starts at the file and
-not at the history.
-
-- [ ] **GPU-texture pooling of cached frames.** The source-frame cache
- (`concat-host/src/pool.rs`) hands out CPU frames; a frame that goes to the
- wgpu compositor is uploaded again each time it is drawn. Pool the textures
- beside the frames now that both compositors read one `FramePlan`.
-- [ ] **Zero-copy hardware decode.** VideoToolbox frames
- (`concat-media/src/hardware.rs`) are copied out to system memory and back
- up to the GPU. An IOSurface handed straight to wgpu skips both copies; the
- seam is the decoder's output frame and the compositor's texture upload.
-- [ ] **Crop, flips and transition fades filled by the export.** `FramePlan`
- carries all three, the parity tests fill them, and the compositors honour
- them - but the export path still bakes them into the decoder's filter
- chain (`concat-export/src/resolve.rs`, `concat-media` filter graph). Fill
- the plan from the export and take them out of the chain.
-- [ ] **Timeline drawing at 2 ms.** Clips are virtualised
- (`TimelinePane::published_span`), but filmstrips and waveforms are still
- one image and one path per clip; the design is one texture per track per
- zoom level. The Slint repaint itself has not been measured
- (`SLINT_DEBUG_PERFORMANCE` needs the window), so the target is not yet
- shown either way.
-- [ ] **OS-native restoration behind Enhance.** Enhance runs restoration
- models through `concat-vision/src/runtime.rs` on every platform. macOS 26
- and iOS 26 carry a super-resolution scaler, frame interpolation and motion
- blur in VideoToolbox (`VTFrameProcessor`; Rust bindings in
- `objc2-video-toolbox`), and Windows App SDK carries a video
- super-resolution API. Both belong behind the same enhanced-copy job as a
- fast path chosen per platform, never as a second feature; the seam is the
- job's per-frame step.
-- [ ] **Shader passes at a size of their own.** A `[[wgsl.pass]]` declares
- `size` over `WIDTH` and `HEIGHT` (`concat-effects/src/manifest.rs`), but
- `run_passes` (`concat-render/src/gpu.rs`) claims every target at the
- source's size and ignores it. Honouring it is what a GPU upscaler package
- (FSR 1.0, Anime4K, both MIT with WGSL ports) needs to write a larger
- picture than it reads.
diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
index 1cfab7b4..52cbb407 100644
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@@ -306,9 +306,31 @@ with; the window's Remote page shows it.
into the decoder's chain, not through the plan.
- `ExportClip` remains the CLI and API wire type and the title rasteriser's
output.
-- Cached frames are uploaded to the GPU per composite; keeping them as
- textures in the pool is the follow-up now that both compositors read a plan.
+- Cached frames are uploaded to the GPU again unless they were drawn in the
+ previous composite (`WgpuCompositor::upload` reuses a texture by frame id);
+ a source-texture cache with its own budget, so a scrub back over cached
+ ground skips the upload too, is the follow-up now that both compositors
+ read a plan. The frame pool is `concat-media/src/pool.rs`.
- Zero-copy hardware frames (IOSurface into wgpu) are not done; a hardware
- frame is transferred to memory first.
-- Filmstrips are one image per clip; one texture per track per zoom level is
- not done, and the Slint repaint itself is not measured headless.
+ frame is transferred to memory first, and the libavfilter stage between
+ the download and the upload (rotation, fit, crop, colour range, RGBA) would
+ have to move to the GPU with it.
+- The preview resolves transitions with fades off, so fade-black, fade-white
+ and the wipes are absent from the monitor until they come through the plan.
+- Filmstrips are one image per media item drawn as up to 120 tile images per
+ clip, and waveforms one path per clip drawn twice; one texture per track
+ per zoom level is not done, and the Slint repaint itself is not measured
+ (`SLINT_DEBUG_PERFORMANCE` needs the window).
+- Enhance runs one restoration model (`concat-vision/src/enhance.rs`) through
+ ONNX Runtime on every platform. The OS scalers (VideoToolbox's
+ `VTFrameProcessor` on macOS 26 and iOS 26, the Windows App SDK's video
+ super-resolution) belong behind the same enhanced-copy job as a per-platform
+ fast path at its per-frame step (`concat-host/src/enhance.rs`, the
+ `enhancer.enhance` call), never as a second feature. Frame interpolation
+ does not fit that step: it changes the frame count and the encoder's rate.
+- A package's `[[wgsl.pass]]` list (`target`, `size` over `WIDTH` and
+ `HEIGHT`; `concat-effects/src/manifest.rs`) is parsed and never read:
+ `run_passes` (`concat-render/src/gpu.rs`) runs one pass per applied effect
+ at the source's size, with no named intermediates. Wiring it up is what a
+ GPU upscaler package (FSR 1.0, Anime4K, both MIT with WGSL ports) needs to
+ write a larger picture than it reads.
diff --git a/AUDIT-2026-09-23-QUICK.md b/AUDIT-2026-09-23-QUICK.md
new file mode 100644
index 00000000..c5bb97b5
--- /dev/null
+++ b/AUDIT-2026-09-23-QUICK.md
@@ -0,0 +1,129 @@
+# Concat audit, 23 Sep 2026 — the quick version
+
+Same audit as `AUDIT-2026-09-23.md`, one screen. Each bar has 10 slots:
+🟢 filled = good, ⚪ empty = missing. Severity: 🔴 critical · 🟠 high · 🟡 medium · ⚪ low.
+
+## 🎯 Verdict
+
+**Structure: done. Guarantees: not yet.** The 16 Sep plan landed in full. What is left is rules the docs promise and the code does not hold yet: bounds, checks, one-place enforcement. Cheaper to fix than what came before.
+
+## 📊 The whole app
+
+| Aspect | Bar | Score |
+|---|---|:-:|
+| 🧹 Code quality | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 |
+| 🏛️ Architecture | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 |
+| 🧭 Design philosophy | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 |
+| 🔧 Maintainability | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 |
+| 😊 User likeability | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 |
+| ⚡ Performance | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 |
+| 📈 Scalability | 🟢🟢🟢🟢⚪⚪⚪⚪⚪⚪ | 4 |
+| 🔐 Security | 🟢🟢🟢🟢⚪⚪⚪⚪⚪⚪ | 4 |
+| 🧪 Testing | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 |
+| 📚 Docs | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 |
+| 🤖 CI and release | 🟢🟢🟢🟢⚪⚪⚪⚪⚪⚪ | 4 |
+
+## 🧱 Per crate (average of its six scales)
+
+| Crate | Bar | Avg | One thing |
+|---|---|:-:|---|
+| concat-core | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 7.3 | 🟢 no deps, rational time · ⚪ linear clip lookup |
+| concat-vision | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.2 | 🟢 ORT fully wrapped · ⚪ serial batch-1 |
+| concat-effects | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.2 | 🟢 strict manifests · ⚪ no host-version or licence field |
+| concat-text | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.0 | 🟢 zero engine deps · ⚪ copies the font per render |
+| concat-api | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.6 | 🟢 every verb JSON, docs match · 🔴 any path, any size |
+| concat-server | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.2 | 🟢 constant-time token · 🔴 no limits at all |
+| concat-media | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.7 | 🟢 unsafe is sound · 🔴 a panicking job wedges the scheduler |
+| concat-project | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.0 | 🟢 commands only · 🟠 ripple copies every clip |
+| concat-speech | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.2 | 🟢 whisper on Metal · 🔴 downloads unverified |
+| concat-export | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5.8 | 🟢 one composite path · ⚪ 2 403-line lib.rs |
+| concat-host | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5.7 | 🟢 small modules · 🔴 corrupt doc opens empty |
+| concat (Rust) | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5.7 | 🟢 panes + echo-and-commit · 🔴 results land in the wrong project |
+| concat (Slint) | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.8 | 🟢 one Modal pattern · ⚪ accessibility 2/10 |
+| concat-render | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5.3 | 🟢 FramePlan · 🟠 7 CPU kernels for 147 shaders |
+
+## 🎛️ Per feature
+
+| Feature | Bar | Score | State |
+|---|---|:-:|---|
+| Launcher | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ |
+| Import and bin | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ |
+| Timeline editing | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ |
+| Undo / redo | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ sharing lost on ripple |
+| Relink | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ |
+| Themes | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | ✅ |
+| Effects and filters | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ⚠️ CPU path never pixel-compared |
+| Titles | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ CPU-heavy |
+| Keyframes | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| Cutout | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| Captions | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| Text-to-speech | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| Templates | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| LUTs | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ |
+| Export | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ serial loop |
+| Hardware decode | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | ✅ not zero-copy |
+| Transitions | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ fades and wipes missing from preview |
+| Crop, blend, masks | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ Darken/Lighten wrong at partial opacity |
+| Enhance | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ one model, no native path |
+| Playback and audio | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ 400 decodes per speed-curved clip |
+| Remote API | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ unsafe defaults, window not a client |
+| Languages | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | ⚠️ 13 of 14 at ~71 %, no RTL |
+| Custom packages | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | ⚠️ trial never runs, multi-pass unwired |
+| Proxies | 🟢🟢🟢🟢⚪⚪⚪⚪⚪⚪ | 4 | ❌ no UI, never swept |
+| Accessibility | 🟢🟢⚪⚪⚪⚪⚪⚪⚪⚪ | 2 | ❌ two accessible-* properties |
+
+## 🚨 Top findings
+
+| # | Sev | What | Where |
+|:-:|:-:|---|---|
+| 0 | 🔴 | **`main` is red**: CI fails at fmt (8 files), 6 tests fail underneath | `ci.yml`, §1.2 of the full report |
+| 1 | 🔴 | Corrupt project opens empty, next save overwrites it | `concat-host/src/session.rs:95` |
+| 2 | 🔴 | Panicking scheduler job wedges the scheduler for good | `concat-media/src/prefetch.rs:355` |
+| 3 | 🔴 | Worker results land in whichever project is open | `concat/src/studio.rs:5376` |
+| 4 | 🔴 | API writes any path; input sizes and line lengths unbounded | `concat-api/src/lib.rs:499,707` · `concat-server/src/json.rs:117` |
+| 5 | 🔴 | Window and Remote page are two engines on one folder | `concat/src/panes/settings.rs:403` |
+| 6 | 🔴 | 20 of 23 model downloads have no digest; 18 URLs unpinned | `models/manifest.toml` |
+| 7 | 🟠 | Hostile-package trial never runs; loop check trivial; no error scope | `concat-render/src/gpu.rs:850,1091` |
+| 8 | 🟠 | Ripple, freeze, fill-slot deep-copy every clip into undo | `concat-project/src/model.rs:1966` |
+| 9 | 🟠 | Stream start time ignored (MTS files offset) | `concat-media/src/decode.rs:743` |
+| 10 | 🟠 | Darken / Lighten maths wrong at partial opacity | `concat-render/src/compositor.rs:258` |
+| 11 | 🟠 | Art scan on every pointer event; full publish 30×/s | `concat/src/lib.rs:373` |
+| 12 | 🟠 | CI is Linux-only until a tag; perf gate never runs; parity skips | `.github/workflows/ci.yml` |
+| 13 | 🟡 | Clamps in four places, they disagree | `concat-project/src/commands/properties.rs:28` |
+| 14 | 🟡 | E2E suite checks timing, not looks | `concat-host/tests/export.rs:915` |
+| 15 | 🟡 | Caches with no bound: proxy, artwork, pinned frames, GPU passes | `proxy.rs`, `prefetch.rs:243`, `gpu.rs:1453` |
+| 16 | 🟡 | 3–5 copies per frame; filter graph rebuilt per treated frame | `decode.rs:934`, `treat.rs:85` |
+| 17 | 🟡 | Hub thread does heavy work in line; gRPC fan-out under a mutex | `hub.rs:52,71` |
+| 18 | 🟡 | ONNX Runtime likely linked twice; pinned to an rc | `src/Cargo.toml:41` |
+| 19 | 🟡 | Speech only reachable from the window | `crates/concat/Cargo.toml` |
+| 20 | 🟡 | Titles copy the whole font per render | `concat-text/src/lib.rs:268` |
+| 21 | ⚪ | Wrong H shortcut, ⌘ on every OS, "twelve" languages, dark colours in light theme | `settings.slint:671`, `tray.slint:127` |
+| 22 | ⚪ | Dead `ui/demo/` still imported | `app.slint:21` |
+| 23 | ⚪ | Stale docs: 429 tests (544), crate map, §7 diagram, mobile.yml, CONTRIBUTING | `ARCHITECTURE.md:276` |
+| 24 | ⚪ | 8 files unformatted; dead `core::Project`; disk IO in a wasm crate | `timeline.rs:500`, `model.rs:1896` |
+
+## 🧾 Today's checks
+
+| Check | Result |
+|---|---|
+| fmt | ⚠️ 8 files would change |
+| clippy | ✅ 0 warnings |
+| tests | ❌ 537 passed, **6 failed**, 1 ignored (3 window, 2 render transitions, 1 Chatterbox) |
+| perf --check | ✅ 14 of 14 within budget (GPU composite 3.2 ms, export 115 fps, hardware decode 326 fps) |
+
+## 🗺️ Do next, in order
+
+0. 🔴 **Make `main` green** — `cargo fmt --all`, then fix the 6 failing tests (5 are today's commits, 1 is locale coverage).
+1. 🔴 **Correctness trio** — not-found vs parse error in `Session::open`; `catch_unwind` around scheduler jobs; a session epoch on worker results.
+2. 🔴 **Server limits + root allowlist** — line cap, pre-auth timeout, connection cap, bounded outbox and sizes; share the window's exporter and sessions.
+3. 🔴 **Verify every model** — fill the digests, pin to commits, reject empty digests in release, generate the tables from the manifest.
+4. 🟠 **Refuse hostile packages for real** — trial at install, error scopes, binding-type check, pipelines keyed by source hash.
+5. 🟠 **Document maths** — `clips_where` in place of `clips_mut`; one clamp pass; `start_time`; fix Darken and Lighten.
+6. 🟡 **Stop the per-event work** — art refresh on change only; targeted publish; cache treat graphs; `av_frame_ref`; sweep proxy and artwork.
+7. 🟡 **CI** — macOS and Windows check; lavapipe with parity required; `perf --check`; release depends on CI; fix the stale docs.
+
+Then the ARCHITECTURE §10 list in its order: plan filled by the export → source textures pooled → zero-copy → one texture per track → native Enhance → multi-pass packages.
+
+## ✅ Keep
+
+Layering · `FramePlan` · commands-only document with Arc undo · hardware decode fallback · package load hygiene · JSON API contract and its docs · pane pattern · model staging and licences · tests that check behaviour.
diff --git a/AUDIT-2026-09-23.md b/AUDIT-2026-09-23.md
new file mode 100644
index 00000000..5bf7b32c
--- /dev/null
+++ b/AUDIT-2026-09-23.md
@@ -0,0 +1,467 @@
+# Concat audit, 23 September 2026
+
+Where the codebase stands one week after the 16 September proposal landed.
+Every crate, every feature and the engineering process were read and scored
+on the same seven scales, plus security, testing and documentation where they
+turned out to matter. Every finding names a file and a line. The sibling
+`AUDIT-2026-09-23-QUICK.md` is the same audit as a one-screen scorecard.
+
+**How it was done.** Nine reviewers each read one part of the tree in full
+(the crates, the window's Rust, the Slint UI, the API and transports, the
+model crates, and the tests, CI, build and docs), scored it and reported
+findings with locations. The headline findings were then re-verified by hand
+against the code before they went in here. Line counts, test counts and the
+unsafe, unwrap and panic tallies come from the tree; `cargo fmt --check`,
+`cargo clippy --workspace --all-targets`, `cargo test --workspace` and
+`cargo run -p concat-perf --release -- --check` were run on this machine
+(Apple Silicon, macOS 26). The app itself was not launched: "user
+likeability" is judged from what the UI code does, not from users.
+
+**Scale.** 0 to 10. 8 and up means "keep doing this"; 5 to 7 means "works,
+with known debt"; 4 and under means "will bite at the next order of
+magnitude, or already bites".
+
+---
+
+## 1. The numbers
+
+### 1.1 Size, tests and hygiene per crate
+
+| Crate | Rust lines | Slint / WGSL | Tests (files) | `unsafe` | `unwrap`/`expect` | `panic!` family |
+|---|---:|---:|---:|---:|---:|---:|
+| concat (window) | 16 058 | 23 636 slint | 40 (11) | 1 | 15 | 1 |
+| concat-project | 9 911 | | 100 (3) | 0 | 312 | 12 |
+| concat-host | 8 727 | | 53 (14) | 1 | 152 | 12 |
+| concat-media | 8 038 | | 79 (11) | 30 | 148 | 2 |
+| concat-render | 4 873 | | 62 (6) | 1 | 39 | 3 |
+| concat-effects | 4 664 | 2 205 wgsl | 46 (7) | 0 | 76 | 2 |
+| concat-export | 3 286 | | 25 (3) | 0 | 17 | 0 |
+| concat-speech | 3 087 | | 19 (3) | 3 | 23 | 0 |
+| concat-vision | 2 760 | | 31 (8) | 0 | 22 | 0 |
+| concat-core | 2 691 | | 45 (7) | 0 | 22 | 2 |
+| concat-api | 1 993 | | 18 (2) | 0 | 19 | 8 |
+| concat-server | 1 183 | | 10 (4) | 0 | 27 | 2 |
+| concat-text | 1 101 | | 11 (1) | 0 | 28 | 1 |
+| concat-perf | 804 | | 2 (1) | 0 | 32 | 0 |
+| concat-cli | 449 | | 3 (1) | 0 | 2 | 0 |
+| concat-android | 335 | | 0 (0) | 10 | 9 | 2 |
+| **Total** | **~71 000** | **~25 800** | **544 (82)** | **46** | **~940** | **47** |
+
+The unwrap and expect counts include test code (concat-project's 312 are
+nearly all in its 3 692-line test module). Outside tests, every crate but the
+window and the host is close to panic-free, and every `unsafe` block in
+concat-media carries a SAFETY note and was found sound.
+
+Other facts: 173 effect packages (147 with WGSL, 20 transitions, 25 audio);
+14 locale files; 606 commits since 25 August, 451 of them this month; 26
+release tags, latest `v0.2.3`; largest files `concat/src/studio.rs` 7 484,
+`concat-export/src/lib.rs` 2 403, `concat-render/src/gpu.rs` 1 794,
+`concat-media/src/decode.rs` 1 603; largest Slint files
+`workspace/media-pane.slint` 2 303 and `inspector/clip-inspector.slint` 2 131.
+
+### 1.2 The checks, run today
+
+| Check | Result |
+|---|---|
+| `cargo fmt --all -- --check` | 8 files would change |
+| `cargo clippy --workspace --all-targets` | 0 warnings |
+| `cargo test --workspace --no-fail-fast` | 537 passed, 6 failed, 1 ignored |
+| `cargo run -p concat-perf --release -- --check` | 14 of 14 scenarios within budget |
+
+**`main` is red.** The last four CI runs on `main` failed at the
+`cargo fmt --check` step, so CI has not run the test suite for today's
+commits. Run locally, six tests fail, five of them in files changed today:
+
+| Test | Crate | What it says |
+|---|---|---|
+| `transitions::tests::a_slide_arrives_from_the_far_side` | concat-render | "the outgoing picture dropped away": red where blue was expected (`transitions.rs:546`) |
+| `transitions::tests::every_shape_runs_from_one_picture_to_the_other` | concat-render | `smoothleft` at progress 1 is a blend, not the incoming picture (`transitions.rs:436`) |
+| `chatterbox::tests::the_text_is_normalised_the_way_resemble_does_it` | concat-speech | a double space survives after a comma (`chatterbox.rs:1100`) |
+| `panes::settings::listen_tests::a_missing_or_bad_half_gets_the_default` | concat | `"192.168.1.9:"` keeps its trailing colon as the host (`settings.rs:753`) |
+| `format::tests::zooming_in_reveals_the_fine_buckets_and_never_loses_a_peak` | concat | every column carries the spike, not one (`format.rs:448`) |
+| `i18n::tests::every_shipped_locale_parses_names_itself_and_keys_off_the_inventory` | concat | 13 locales lack about 290 of the 1 015 keys (`i18n.rs:254`) |
+
+The first five look like today's work outrunning its tests; the last is a
+rule (every locale covers the inventory) the locale files have not met since
+the inventory grew. The end-to-end export suite, the parity fixtures and the
+migration tests all pass.
+
+**The perf table**, release build, every scenario inside its budget:
+
+| Scenario | Measured | Budget |
+|---|---:|---:|
+| plan one frame of 200 clips on 4 tracks | 0.53 µs | ≤ 200 µs |
+| 200 clip additions and 200 undos | 0.41 ms | ≤ 150 ms |
+| ripple delete of 200 clips out of 400 | 0.14 ms | ≤ 20 ms |
+| save and open a 200-clip document | 1.49 ms | ≤ 60 ms |
+| choose the visible clips of 200 for one publish | 0.20 µs | ≤ 50 µs |
+| read an hour's waveform across 2048 columns | 4.33 µs | ≤ 2000 µs |
+| decode 1080p h264 in software | 1 637 fps | ≥ 120 |
+| decode 1080p h264 on the hardware (VideoToolbox) | 326 fps | ≥ 120 |
+| scrub forward over 90 frames, cold | 0.81 ms/frame | ≤ 25 |
+| scrub back over covered ground | 0.24 µs/frame | ≤ 200 |
+| decodes when an effect knob is turned | 0 | ≤ 0 |
+| compose 1080p, 3 layers, sepia, CPU | 54.7 ms | ≤ 400 |
+| compose 1080p, 3 layers, sepia, GPU | 3.19 ms | ≤ 25 |
+| export a 4 s 720p cut of two overlapping clips | 115 fps | ≥ 30 |
+
+Two readings worth keeping in mind: hardware decode is five times slower
+than software here, because the frame is downloaded and converted on the CPU
+after the chip decodes it (finding 16 and ARCHITECTURE §10), and the CPU
+composite is seventeen times the GPU one, which is what a machine without a
+GPU, and any export whose device dies mid-run, will see.
+
+---
+
+## 2. Scorecard for the whole app
+
+| Aspect | Score | One line |
+|---|:---:|---|
+| Code quality | 7 | Clippy-clean, typed errors, sound and documented `unsafe`; long functions and copied blocks pull it down. |
+| Design architecture | 7 | The crate layering holds and FFmpeg and ONNX Runtime are contained; the window bypasses `Session` and the API, and chains bypass `FramePlan`. |
+| Design philosophy | 6 | The ideas are clear and written down. Several are stated and not enforced: one clamp place, hostile packages refused, publish by diff, stale results dropped in one place. |
+| Maintainability | 6 | Excellent comments and one shape per pane; four files over 1 800 lines and model tables kept in five places. |
+| User likeability | 7 | A broad CapCut-class feature set with real polish; accessibility is near zero and a few UX facts are stale. |
+| Performance | 5 | The planner and undo are microseconds; the per-frame copies, the UI's per-event art scan and the serial export loop are not. |
+| Scalability | 4 | Linear clip lookups, undo that copies every clip on a ripple, caches with no bound, one server thread, unbounded socket input. |
+| Security | 4 | The token is done right. A token holder can write any path; input sizes are unbounded; 20 of 23 model downloads have no digest. |
+| Testing | 6 | 544 behaviour tests and a real end-to-end suite; it checks timing, not looks, and the GPU parity suite silently skips in CI. |
+| Documentation | 7 | Substantial and well written, with a handful of stale claims. |
+| CI and release | 4 | `main` is red at the fmt step; Linux only until a tag; the perf gate the docs describe does not run. |
+
+---
+
+## 3. Scores per crate
+
+Columns: code quality, maintainability, scalability, performance,
+architecture, design philosophy. A dash means the scale does not apply.
+
+| Crate | CQ | Maint | Scale | Perf | Arch | Phil | Note |
+|---|:-:|:-:|:-:|:-:|:-:|:-:|---|
+| concat-core | 8 | 8 | 6 | 6 | 8 | 8 | No deps, rational everywhere; linear `clip_on_track_at`, gcd on every add. |
+| concat-project | 7 | 6 | 5 | 5 | 7 | 6 | Commands are the one way to write; clamps live in four places; ripple copies every clip. |
+| concat-media | 7 | 7 | 5 | 6 | 8 | 7 | Unsafe is sound; 3 to 5 copies per frame; scheduler cannot survive a panicking job. |
+| concat-render | 7 | 5 | 4 | 5 | 6 | 5 | `FramePlan` is excellent; 7 CPU kernels for 147 shaders; GPU memory unbounded at 8K. |
+| concat-effects | 8 | 8 | 6 | 7 | 7 | 7 | Strict manifests and a shared WGSL prelude; format lacks host version, capabilities, licence. |
+| concat-export | 6 | 4 | 6 | 5 | 7 | 7 | One composite path for export and preview; a 2 403-line lib.rs and 12 side tables per clip. |
+| concat-host | 6 | 6 | 5 | 6 | 5 | 6 | Small clear modules; `Session::open` swallows parse errors; proxy and artwork caches never swept. |
+| concat (Rust) | 7 | 5 | 5 | 5 | 6 | 6 | Panes and echo-and-commit are sound; 67 callbacks skip the `Msg` loop; art scan per event. |
+| concat (Slint) | 7 | 6 | – | – | 7 | 7 | One `Modal` pattern and one token file; accessibility 2, i18n 6. |
+| concat-api | 8 | 8 | 5 | – | 8 | 9 | Every verb JSON, docs match; security 3: any path, any size. |
+| concat-server | 7 | 8 | 4 | – | 8 | 9 | Constant-time token, loopback default; no limits, one Hub thread; security 5. |
+| concat-cli | 7 | 8 | – | – | 7 | – | `serve` parks forever, so a Unix socket file is left behind. |
+| concat-android | 8 | 7 | – | – | – | – | Justified `unsafe`; no tests. |
+| concat-speech | 7 | 5 | 6 | 7 | 5 | 7 | Whisper on Metal; models listed in five places; only the window can reach it; security 3. |
+| concat-vision | 8 | 7 | 5 | 6 | 9 | 8 | ORT fully wrapped; batch-1 serial inference under one lock; security 6. |
+| concat-text | 8 | 8 | 5 | 4 | 9 | 8 | No engine deps; copies the whole font file per render, draws on a frame-sized canvas. |
+
+---
+
+## 4. Scores per feature
+
+Status: **done**, **partial** (works with a known gap), **thin** (present, weak).
+
+| Feature | Where | Status | Score | Why |
+|---|---|---|:-:|---|
+| Launcher, new project, recents | `ui/start.slint` | done | 8 | Three shapes by width; the one screen with a phone layout. |
+| Import, bin, drag and drop | `ui/workspace/media-pane.slint` | done | 8 | Filters, sort, probe on a worker. |
+| Timeline editing | `ui/timeline/`, `panes/timeline.rs` | done | 8 | Split, ripple, snap, several timelines, virtualised clips. |
+| Undo and redo | `concat-project/src/editor.rs` | done | 8 | Arc-shared snapshots, gestures as one step; sharing lost on ripple (finding 8). |
+| Transitions | `concat-export/src/lib.rs:467` | partial | 6 | Fades and wipes baked into chains; absent from the preview. |
+| Effects and filters | `concat-effects/packages/` | partial | 7 | 173 packages; CPU export renders them through FFmpeg chains never pixel-compared with the WGSL. |
+| Titles and text | `concat-text` | done | 7 | Correct; CPU-heavy per animated frame. |
+| Keyframes, animations | `concat-project/src/model.rs:1573` | done | 7 | Keys survive split, trim and merge; offset and rotation keys unclamped. |
+| Crop, transform, blend, masks | `concat-render/src/plan.rs` | partial | 6 | Darken and Lighten wrong at partial opacity on both paths (finding 10). |
+| Cutout, smart brush | `concat-vision/src/segment.rs` | done | 7 | Compiled-in fallback model; serial batch-1 inference. |
+| Enhance | `concat-host/src/enhance.rs` | thin | 6 | One Real-ESRGAN model on every platform; native scalers not used. |
+| Auto-captions | `ui/dialogs/captions.slint` | done | 7 | Whisper on Metal, with progress; results can land in the wrong project (finding 3). |
+| Text-to-speech, voice from a sample | `concat-speech/src/tts.rs` | done | 7 | Kokoro, Pocket, Chatterbox; downloads unverified (finding 6). |
+| Templates | `concat-host/src/templates.rs` | done | 7 | |
+| LUT import | `concat-effects/src/cube.rs` | done | 7 | Tables over 65 a side refused. |
+| Custom and community packages | `concat-effects/src/catalogue.rs` | thin | 5 | Multi-pass parsed and never run; the trial never runs before enabling (finding 7). |
+| Proxies | `concat-host/src/proxy.rs` | thin | 4 | No UI, no sweep of `cache/proxy`. |
+| Export | `ui/dialogs/export.slint`, `concat-export` | done | 7 | Quality tiers, 10-bit, CBR and VBR; serial loop; one social preset. |
+| Playback and audio | `concat-host/src/playback.rs` | partial | 6 | A speed-curved clip becomes up to 400 decodes; the mixer scans every clip per sample. |
+| Hardware decode | `concat-media/src/hardware.rs` | done | 7 | Sound, falls back once; frames still round-trip through memory. |
+| Remote API | `concat-api`, `concat-server` | partial | 6 | Clean contract; unsafe defaults (finding 4); the window does not use it (finding 5). |
+| Relink missing media | `ui/dialogs/relink.slint` | done | 8 | |
+| Themes and accent colours | `ui/theme/` | done | 8 | Two lane colours hard-coded to dark. |
+| Languages | `concat/locales/` | partial | 6 | 14 files, 13 at about 71 % of the 1 015 keys; Persian ships without right-to-left. |
+| Accessibility | the whole UI | thin | 2 | Two `accessible-*` properties in 23 636 lines; 12 px base type. |
+
+---
+
+## 5. Findings, by severity
+
+Each carries the location and, where I re-verified it by hand, **(verified)**.
+
+### Critical
+
+1. **A corrupt document opens as an empty project, and the next save
+ overwrites it.** (verified) `Session::open` matches `Err(_)` from
+ `projects::read_document` and builds an empty editor
+ (`concat-host/src/session.rs:95`); that function returns the same `Err` for
+ "file not found" and for a JSON parse failure (`projects.rs:211-215`). The
+ comment nine lines above says a corrupt document is an error. Saves are
+ atomic with fsync (`projects.rs:168-190`), so the window is narrow: a power
+ cut, a sync tool or an external edit. When it happens it is silent data
+ loss.
+
+2. **A panicking scheduler job wedges the scheduler.** (verified)
+ `prefetch.rs:355` runs `job()` with no unwind guard. A panic skips the
+ decrements at 357-363, `running` never drops, `drain` waits forever
+ (`prefetch.rs:307`) and background lanes stay blocked once the counters sit
+ at their limit. Failed thread spawns are also ignored (`prefetch.rs:194`).
+
+3. **Worker results land in whichever project is open.** `Studio::handle`
+ (`concat/src/studio.rs:5376-5429`) checks no project identity, and
+ `close_project` (`studio.rs:5308`) does not reset the export, captions or
+ speech panes. A probe, a caption run or a speech job started in project A
+ delivers into project B (`panes/media_bin.rs:153`, `panes/captions.rs:132`,
+ `panes/speech.rs:339`). ARCHITECTURE §5 claims the opposite.
+
+4. **The API lets a token holder write anywhere and send anything.**
+ (verified) `preview.frame` creates folders and truncates any path
+ (`concat-api/src/lib.rs:707-714`); `export.run`, `project.create` and
+ `template.instantiate` take any path. Width and height are only checked
+ for zero (`lib.rs:499-502`), so 65535×65535 asks for 17 GB on the Hub
+ thread. The JSON transport reads lines of unbounded length before
+ authentication, spawns two threads per connection with no cap and no
+ timeout, and queues output in an unbounded channel
+ (`concat-server/src/json.rs:63,117,193`). The token itself is right:
+ 128 bits, constant-time compare, required on loopback.
+
+5. **The window and the Remote page are two engines on one folder.**
+ (verified) Settings › Remote starts `Server::start(config, Api::new)`
+ (`concat/src/panes/settings.rs:403`): its own session registry and its own
+ exporter. An edit from either side is overwritten by whoever saves last,
+ and "one export at a time" no longer holds across the two.
+ `Api::exporter()` exists for sharing and nothing calls it. The §7 diagram
+ showing the window over the API is wrong.
+
+6. **Model downloads are unverified and unpinned.** (verified) 20 of 23 rows
+ in `models/manifest.toml` have an empty `sha256`; `verify` accepts an empty
+ digest (`concat-host/src/models.rs:289-291`); 18 upstream URLs point at
+ `resolve/main`. The source list includes a mirror and a user-set base URL
+ (`models.rs:149-180`), and a dropped transfer resumes with a Range request
+ against the *next* source (`models.rs:216-219`), relying on the digest to
+ catch a splice.
+
+### High
+
+7. **A hostile package is not really refused.** (verified that `trial` has
+ no production caller) `WgpuCompositor::trial` is only called from tests.
+ The loop check asks only whether a `break` token exists
+ (`concat-effects/src/shader.rs:507-549`); binding types are never checked;
+ validation uses `Capabilities::all()` (`shader.rs:614`). A pipeline that
+ fails to build does so outside any error scope (`gpu.rs:1091`), which in
+ wgpu is a panic.
+
+8. **Undo's structural sharing is defeated by every ripple.** (verified)
+ `clips_mut()` calls `Arc::make_mut` on every clip before the caller filters
+ (`concat-project/src/model.rs:1966`); ripple trim, ripple delete, freeze
+ and fill-slot use it (`commands/clips.rs:245,468,596,653`,
+ `commands/media.rs:105`). The only sharing test covers `UpdateClip`.
+
+9. **The stream's start time is ignored.** (verified: no reference in the
+ crate) Timestamps are used raw (`decode.rs:743`) and turned straight into
+ frame indexes (`pool.rs:364`). MTS and other camera files that start at a
+ non-zero time will be offset in the pool, the paced export and the
+ duration clamp. Inferred from the code, not reproduced.
+
+10. **Darken and Lighten are wrong at partial opacity, on both paths.**
+ (verified) Both compute `min(colour·α, ground)`
+ (`compositor.rs:258-259`, `gpu.rs:421-430`): a white Darken layer at 30 %
+ turns a mid-grey ground dark. Parity cannot catch it because both agree.
+
+11. **The UI thread scans every media item on every pointer event.**
+ (verified) The `handler!` macro calls `refresh_art()` after each callback
+ (`concat/src/lib.rs:373`), including the lanes' hover and drag stream.
+ That loops over all media calling `proxy::ensure` (two `stat` calls each,
+ `studio.rs:2010-2030`) and decodes cached JPEGs synchronously. Separately,
+ playback runs a full `publish_chrome` about 30 times a second
+ (`host.rs:197`, `studio.rs:1905`), rebuilding every bin row, both menus and
+ the recents list each time.
+
+12. **`main` is red, and CI never compiles macOS or Windows until a tag.**
+ Today's four CI runs on `main` fail at `cargo fmt --check` (8 files),
+ which hides six failing tests behind it (§1.2). `build-app.yml` is
+ `workflow_call` only; `ci.yml` runs on `ubuntu-latest`; `release.yml`
+ does not require CI to pass. The GPU parity suite returns early with no
+ adapter (`gpu/tests.rs:20-23`) and the runner installs none, so the
+ parity row in ARCHITECTURE §8 is enforced only on developers' machines.
+ No workflow runs `concat-perf --check`; 9 of its 13 scenarios never run
+ in CI.
+
+### Medium
+
+13. **Clamps live in four places and disagree.** Commands re-implement
+ `Clip::tidy` (`commands/properties.rs:28-37,340-355`); `TextStyle::tidy`
+ runs only on load (`model.rs:1418`); the minimum transition length is
+ applied only in tidy; `SetClipKey` leaves offset and rotation unclamped
+ (`properties.rs:198`). A document can change on save and reopen. The
+ speed range is written three times (`model.rs:1090`,
+ `concat-core/src/retime.rs:27`, `concat-media/src/audio.rs:35`).
+
+14. **The end-to-end suite checks timing, not looks.** It samples the centre
+ pixel and maps colour to a source second (`tests/export.rs:305-323`),
+ which proves cuts, trims, speed and reverse. Transitions, effects, keys,
+ transform, crop, blend, fades, titles and adjustment layers assert only
+ length and sound (`export.rs:915-1064`).
+
+15. **Caches with no bound.** `cache/proxy` and artwork are never swept
+ (`proxy.rs:16` says otherwise; `media.rs:323`); frames pinned by the
+ prefetcher are neither deduplicated nor capped and `Monitor::prefetch`
+ keeps adding while paused (`prefetch.rs:243,266`, `preview.rs:255`); the
+ 8 warm readers each decode with every core (`pool.rs:531`,
+ `decode.rs:651`) outside the byte budget; on the GPU, `run_passes` claims a
+ fresh full-size target per pass (`gpu.rs:1453`, about 133 MB each at 8K)
+ and the LUT and reveal caches are never evicted (`gpu.rs:198-201`).
+
+16. **Per-frame copies and graph rebuilds.** Every paced frame is deep-copied
+ before filtering (`decode.rs:934,944,994`), then filtered, then copied row
+ by row into a `Frame` (`decode.rs:873`); a hardware frame adds a download.
+ On the monitor path every treated frame builds, parses and validates a
+ fresh filter graph (`treat.rs:85-126`).
+
+17. **The Hub thread does heavy work in line, and event fan-out holds a
+ mutex.** Preview decode and composite, probes and title painting run on
+ the one thread that serves every caller (`hub.rs:71`, `lib.rs:505,396`).
+ The gRPC subscriber does a `blocking_send` into a channel of 64 while the
+ subscribers mutex is held (`grpc.rs:70-74`, `hub.rs:52`): one slow client
+ stalls the export thread.
+
+18. **ONNX Runtime is probably linked twice** (sherpa-onnx's static copy plus
+ `ort` 2.0.0-rc.13 through vision and Chatterbox), and CoreML is registered
+ with default options and an aspect-dependent input shape
+ (`concat-vision/src/runtime.rs:141`, `segment.rs:115`). Needs profiling.
+
+19. **Speech is reachable only from the window.** Only `concat` depends on
+ `concat-speech`, so captions and TTS have no API, CLI or server verb;
+ "three ways in" does not hold for them. Also missing verbs: enhance,
+ proxies, package install, cache clear, a jobs list, CBR export.
+
+20. **The title rasteriser copies the whole font file per render** and draws
+ on a frame-sized canvas it then un-premultiplies pixel by pixel
+ (`concat-text/src/lib.rs:268-279,529-539`). With a CJK `.ttc` that is tens
+ of megabytes per animated frame.
+
+### Low
+
+21. **UI facts that are wrong or stale.** Settings says H is "Flip
+ horizontally" (`ui/dialogs/settings.slint:671`) while H is the pan tool
+ (`app.slint:418`); it lists 3 of about 27 bindings. Tooltips show ⌘ on
+ every platform (`ui/timeline/tray.slint:127,134`). README says twelve
+ languages; there are 14 files. Two lane colours ignore the theme
+ (`lanes.slint:792,835`). The window minimum is 900×560 while Android is
+ listed as supported.
+
+22. **Dead Slint still imported.** `ui/demo/` (495 lines) is unreachable, yet
+ `EffectData` is imported from it by `app.slint:21`, `editor.slint:4` and
+ `workspace/config-pane.slint:9`.
+
+23. **Stale docs.** ARCHITECTURE §8 says 429 tests in 75 files (544 in 82);
+ §1's crate map omits speech → media; §7's diagram puts the window over the
+ API; `src/README.md:115` says `mobile.yml` runs on push (it does not);
+ `release.yml:13` describes archives the build does not ship;
+ `CONTRIBUTING.md:43` calls `test/` fixtures (it is Python prototypes).
+ Newcomer setup omits ONNX Runtime, libclang, ALSA and GTK.
+
+24. **Small hygiene.** 8 files fail `cargo fmt --check`. `core::Project`
+ (`timeline.rs:500`) is dead. `missing_media` does disk IO inside a crate
+ that must build for wasm (`model.rs:1896`). The CLI's `serve` parks
+ forever and never removes its Unix socket (`concat-cli/src/main.rs:226`).
+ `lib.rs:23` in the window uses `slint::private_unstable_api`. Pipelines
+ are keyed by `id@version`, so a WGSL edit without a version bump keeps the
+ stale pipeline (`shader.rs:417`).
+
+---
+
+## 6. What is good, and should stay that way
+
+- **The layering.** Every arrow points one way. FFmpeg is only in
+ concat-media; ONNX Runtime is only behind `concat-vision/src/runtime.rs`;
+ concat-core has no dependencies and builds for wasm.
+- **`FramePlan`.** Geometry and shading are computed once for both
+ compositors (`plan.rs:140-416`), so placement and weighting cannot drift.
+- **The document.** Commands are the only write path; snapshots share through
+ Arc; gestures fold into one step; no-ops record nothing; batches are
+ all-or-nothing; unknown fields survive a round trip; migrations are a table.
+- **Hardware decode.** Reference-counted device, `get_format` fallback, a
+ failed download reopens in software and seeks back to the exact frame,
+ with tests that the frame count and timing match.
+- **Package hygiene at load.** `deny_unknown_fields`, every chain rendered at
+ its defaults, minimum and maximum, an FFmpeg filter allowlist, LUT paths
+ held to the folder, the `concat.` namespace reserved.
+- **The API contract.** Every verb is JSON named `area.verb`; gRPC wraps the
+ same JSON; the docs match the code on every method spot-checked.
+- **The window's pane pattern.** State + `Msg` + `update` + `data`, with the
+ pane taken out during `update`; the monitor keeps one request in flight and
+ builds its texture on the event loop on purpose.
+- **Model fetching.** `.part` files, staging, rename into place, a
+ compiled-in fallback cutout model, licences recorded for every model.
+- **Tests that check behaviour.** Synthetic media coloured by source second;
+ SSIM parity; migration tests against v1 documents; token handshake tests
+ over TCP and gRPC.
+
+---
+
+## 7. Where we stand
+
+The 16 September proposal had eleven steps. All eleven landed on `main`
+within the week: the bug-fix pass, structural-sharing undo, `Clip::blank` and
+`Clip::tidy`, `resolve.rs`, the tolerant reader with migrations, keys that
+stay on their instant, ten panes behind one `Msg` loop, the source cache and
+scheduler, `FramePlan` with parity, hardware decode, timeline virtualisation
+and the loopback token. The tree is now the shape the proposal drew.
+
+What that week did not buy is enforcement. Almost every serious finding here
+is a stated rule the code does not yet hold: a corrupt document is an error
+(it is not), a hostile package is refused (it is not tried), clamps live in
+one place (four), stale results are dropped in one place (nowhere), publish
+by diff (rows are rebuilt), one engine handle (two). The debt has moved from
+structure to bounds and guarantees, which is the right direction, and is
+cheaper to pay.
+
+**Recommended order.** Each line is a day or a few, and each closes a
+finding above.
+
+1. Correctness trio: `Session::open` distinguishes not-found from parse
+ error; `catch_unwind` with a counter guard around scheduler jobs; a
+ session epoch captured in `spawn` and checked in `deliver`. (1, 2, 3)
+2. Server limits and a root allowlist: line length cap, pre-auth timeout,
+ connection cap, bounded outbox, bounded preview and export sizes,
+ canonical-path check against configured roots. Hand the Remote server the
+ window's exporter and sessions. (4, 5)
+3. Fill every `sha256`, pin upstreams to commits, reject an empty digest in
+ release builds, generate the five model tables from the manifest. (6)
+4. Run `trial` at install at a realistic size, wrap pipeline creation in an
+ error scope that disables the package, check binding types, key pipelines
+ by source hash. (7)
+5. `clips_where(pred)` in place of `clips_mut`, with `ptr_eq` tests for
+ ripple and freeze; one clamp pass after every command; subtract
+ `start_time` in one place; fix Darken and Lighten as
+ `mix(ground, min(colour, ground), α)`. (8, 9, 10, 13)
+6. Art refresh only on revision change; targeted publish from `deliver`;
+ cache treat graphs by spec and size; `av_frame_ref` instead of deep
+ clone; sweep proxy and artwork. (11, 15, 16)
+7. CI: macOS and Windows `cargo check` plus engine tests; lavapipe on Linux
+ with parity required; `concat-perf --check` as a step; release depends on
+ CI. Fix the stale docs in the same pass. (12, 23)
+
+After those, the architectural follow-ups in ARCHITECTURE §10 stay in their
+order: the plan filled by the export, source textures pooled, zero-copy
+hardware frames, one texture per track, native scalers behind Enhance,
+multi-pass packages.
+
+---
+
+## 8. Caveats
+
+Scores are judgements over code, calibrated across reviewers but not
+measured against users or against another editor. Performance findings are
+from reading hot paths plus the perf table; the window's repaint was not
+measured. The OS-side claims in §10 of ARCHITECTURE (macOS 26 scalers, the
+Windows super-resolution API) were not checked. Findings marked (verified)
+were re-read by hand; the rest rest on one reviewer's reading and a line
+number you can open.