Skip to content

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

Create AUDIT-2026-10-04.md

Jareer0895bf096df6
diff --git a/audit/AUDIT-2026-10-04.md b/audit/AUDIT-2026-10-04.md
new file mode 100644
index 00000000..82974c62
--- /dev/null
+++ b/audit/AUDIT-2026-10-04.md
@@ -0,0 +1,759 @@
+# Concat audit, 4 October 2026
+
+Where the codebase stands six days after the 28 September audit, at
+`a1e8f11` (3 October), 60 commits past the 0.2.5 release. Same method, same
+seven scales, same nine reviewers' split, and every one of the 18 findings
+from the 28th (and the handful carried from the 23rd) re-verified against
+today's tree. Every finding names a file and a line. The sibling
+`AUDIT-2026-10-04-QUICK.md` is this audit as a one-screen scorecard.
+
+**How it was done.** Nine reviewers each read one part of the tree in full
+(core and project; media; render and effects; export and host; the window's
+Rust; the Slint UI and locales; the API, server and CLI; speech, vision,
+text, Android and the model manifest; tests, CI, build, docs and release),
+re-verified the 28 September findings assigned to them, scored their part
+and reported new findings with locations. The headline items (the
+`[[replaces]]` hijack, the playback audio offset, the red CI) were re-read
+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.
+
+**Status of this document.** Complete: all nine reviews are in and folded
+in below. No row is provisional.
+
+**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
+
+Deltas are against the 28 September table. The unwrap column gives the
+total and, in brackets, the count outside test modules and test files; the
+panic column is the non-test count.
+
+| Crate | Rust lines | Δ | Slint / WGSL | Tests (files) | `unsafe` | `unwrap`/`expect` (non-test) | `panic!` family (non-test) |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| concat (window) | 21 261 | +1 697 | 28 796 slint | 64 (16) | 12 | 30 (6) | 0 |
+| concat-host | 11 869 | +329 | | 79 (17) | 1 | 230 (2) | 0 |
+| concat-project | 9 465 | +273 | | 99 (1) | 0 | 319 (11) | 6 |
+| concat-media | 9 494 | +453 | | 90 (11) | 37 | 181 (23) | 0 |
+| concat-render | 8 124 | +80 | shaders inline in `gpu.rs` | 88 (6) | 1 | 90 (8) | 4 |
+| concat-effects | 7 096 | +6 | 1 764 wgsl | 69 (9) | 0 | 129 (10) | 2 |
+| concat-export | 3 702 | +204 | | 29 (3) | 0 | 24 (0) | 0 |
+| concat-speech | 3 109 | +15 | | 19 (3) | 3 | 23 (2) | 0 |
+| concat-core | 2 854 | +25 | | 48 (7) | 0 | 26 (4) | 2 |
+| concat-vision | 2 760 | 0 | | 31 (8) | 0 | 22 (2) | 0 |
+| concat-api | 2 351 | +1 | | 21 (2) | 0 | 23 (0) | 0 |
+| concat-server | 1 376 | 0 | | 12 (4) | 0 | 29 (2) | 0 |
+| concat-text | 1 974 | +734 | | 22 (1) | 1 | 53 (1) | 0 |
+| concat-perf | 1 092 | 0 | | 2 (1) | 0 | 34 (34) | 0 |
+| concat-cli | 514 | 0 | | 3 (1) | 0 | 2 (2) | 0 |
+| concat-android | 335 | 0 | | 0 (0) | 10 | 9 (9) | 2 |
+| **Total** | **87 376** | **+3 800** | **~30 600** | **676 (90)** | **65** | **~1 200 (116)** | **16** |
+
+The tree grew by a twentieth in six days, almost all of it in the window
+(`concat/src` +4 148 / −917 and the UI +1 291 / −168) and in concat-text,
+which took on shapes and grew by three fifths. `studio.rs` is now 10 029
+lines (9 461 on the 28th; 47 % of the window's Rust) and was not split.
+The engine crates barely moved: concat-api gained one line, concat-server,
+concat-cli, concat-vision and concat-perf none. Twelve new `unsafe` sites
+are the Wayland drop target (`concat/src/wayland_drop.rs`, 9) and the Linux
+window edges, every one with a SAFETY line.
+
+Other facts: 123 effect packages (98 on format 2 with WGSL: 21 effects, 57
+looks, 20 transitions, with 344 probe tables between them; 25 audio
+packages on FFmpeg chains, by design); 14 locale files at 1 001 keys each,
+0 missing and 0 stale per `scripts/locales.py --check` (but see finding 4);
+801 commits since 25 August, 60 of them since the last audit (509 files,
++10 700 / −1 600; all 60 by the maintainer, no merges); 29 tags, latest
+`v0.2.5`, 71 commits behind `main`; largest files `concat/src/studio.rs`
+10 029, `concat-project/src/lib.rs` 3 851 (3 800 of them tests),
+`concat-render/src/gpu.rs` 3 630, `concat-export/src/lib.rs` 2 873,
+`concat-effects/src/shader.rs` 2 412; largest Slint files
+`workspace/media-pane.slint` 2 425 and `inspector/clip-inspector.slint`
+1 959. `Cargo.lock` holds 817 packages, 63 crate names at two or more
+versions.
+
+### 1.2 The checks, run today
+
+| Check | Result |
+|---|---|
+| `cargo fmt --all -- --check` | **10 files would change, 22 sites** (0 on the 28th) |
+| `cargo clippy --workspace --all-targets` | 0 warnings |
+| `cargo test --workspace --no-fail-fast` | TEST_RESULT |
+| `cargo run -p concat-perf --release -- --check` | PERF_RESULT |
+| CI on `main` | **red at head, and at every run since the 28th** |
+
+**fmt is red, and so is CI.** The ten files are `concat/src/studio.rs`,
+`concat-export/src/lib.rs`, `concat-host/src/dirs.rs` and `titles.rs`,
+`concat-media/src/encode.rs`, `concat-project/src/lib.rs`,
+`commands/mod.rs` and `commands/clips.rs`, `concat-render/src/gpu.rs` and
+`concat-text/src/lib.rs`. The first three sites date from the afternoon of
+the 28th (`35b80d1`, `8d3bf19`, `dc12b64`), after the last audit was
+written; the rest came with `f97c745`, `0699b44` and `21855a7`. The CI
+engine job runs `cargo fmt --check` first (`.github/workflows/ci.yml:93`)
+and has failed there on every completed run since `a262553` (28 September
+22:16 UTC): 11 failures, 29 cancelled by the next push, 0 successes. The
+last green CI run is `b78ee56`, the 0.2.5 release commit. Clippy, the
+locale and model inventories, the 674 tests, the quick perf gate and the
+gRPC build (`ci.yml:98-126`) have not executed on `main` for 71 commits.
+The Nix job, which only builds, is green as of 2 October. The release
+workflow needs CI (`release.yml:66-68`), so a 0.2.6 tag today would produce
+no bundles. Locally, clippy is clean.
+
+TEST_PARAGRAPH
+
+PERF_TABLE
+
+---
+
+## 2. Scorecard for the whole app
+
+| Aspect | 28 Sept | Today | One line |
+|---|:-:|:-:|---|
+| Code quality | 7 | 6 ↓ | Clippy-clean; but fmt is red in 10 files, 6 UI strings bypass i18n, 28 colours are named outside `theme/`, and `studio.rs` grew another 568 lines. |
+| Design architecture | 8 | 8 | Held: one compositor, one plan, the lent device now survives a loss and recovers on a fresh or software adapter; the hold for a dissolve is one flag the renderers agree on. |
+| Design philosophy | 7 | 6 ↓ | Six stated rules not held: "every operation the window can perform is one an API caller can", "nothing outside theme/ names a colour", "every string goes through I18n", "no button below 44 px", "one definition of source time", and ARCHITECTURE's hardware-decode switch that no longer exists. |
+| Maintainability | 6 | 5 ↓ | `studio.rs` 10 029 lines, `concat-export/src/lib.rs` 2 873 with a 41-field `ExportClip`, `ui/demo/` still compiled in for one struct, `tts.rs` 1 244 lines across three model families. |
+| User likeability | 7 | 7 | A razor, a scrollbar, a volume line, keyboard menus, Save Frame As, a scrub strip in the export sheet; against that an export that defaults to 30 fps on a 25 fps cut, eight locales a fifth English, and no dialog a keyboard can press. |
+| Performance | 6 | 6 | Decoded frames go through two CPU copies (three on the chip) before upload; a full publish per monitor frame during playback; the export loop unchanged. |
+| Scalability | 5 | 5 | Art caches never released across projects; the fd leak per API connection, the one Hub thread and the four unswept cache kinds all as they were. |
+| Security | 5 | 4 ↓ | The update trust chain unchanged; new: a user package's `[[replaces]]` can hijack every link to a live built-in at open, and a timed-out install trial kills the monitor's compositor. |
+| Testing | 7 | 6 ↓ | 676 tests, +26 on +3 800 lines; no CI test run for 71 commits; the E2E suite untouched while transitions, titles, shapes and logs gained code; the two locale-flaky tests still race. |
+| Documentation | 7 | 6 ↓ | No Unreleased section for ~20 user-visible changes; the README sells an MCP transport that does not exist; three ARCHITECTURE sentences describe code that changed. |
+| CI and release | 6 | 3 ↓ | Zero green runs since 0.2.5; the gate fired on every push and was ignored; CI builds on `stable` (1.99) while `rust-toolchain.toml` pins 1.93; signing still only when secrets are set. |
+
+---
+
+## 3. Scores per crate
+
+Columns: code quality, maintainability, scalability, performance,
+architecture, design philosophy, security, testing. A dash means the scale
+does not apply. Arrows are against 28 September. Every row rests on a
+full read this time; nothing is provisional.
+
+| Crate | CQ | Maint | Scale | Perf | Arch | Phil | Sec | Test | Note |
+|---|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|---|
+| concat-core | 8 | 8 | 6 | 6 | 8 | 8 | – | 7 | Held. `hold` landed as one flag read in two places with a test; dead `Project`, linear `clip_on_track_at` and a caller-less `source_duration` remain. |
+| concat-project | 8 | 7 | 6 | 6 | 7 | 7 | – | 7 | Held. Shapes copied the title pattern cleanly through settle, tidy and `has_non_finite`; but a shape cannot be edited once placed, an unknown kind is rewritten as a square, and 0.2.5 deletes them. |
+| concat-media | 7 | 7 | 5 | 7 ↑ | 8 | 7 ↓ | 6 | 7 | Copy count known now (2 CPU copies, 3 on the chip); every `unsafe` carries SAFETY; playback audio ignores the stream start while the export mix honours it (finding 3). |
+| concat-render | 7 | 5 | 6 | 7 | 8 | 7 | 6 | 7 | Parity for 10 plan features, 344 probes, device loss recovered; the install trial runs on the monitor's compositor and kills it on a timeout. |
+| concat-effects | 8 | 8 | 6 | 8 | 8 | 7 ↓ | 6 | 8 | `matches_category` hard-codes nine ids against "effects are data"; `[[replaces]]` unfenced (finding 2); `Package::replacing` still drops keys. |
+| concat-export | 6 | 4 | 6 | 5 | 8 | 7 | 6 | 7 | Device loss handled in three steps with the frame redrawn; `lib.rs` 2 873 lines, 41-field `ExportClip`, serial loop; the hold plays keyed effects early. |
+| concat-host | 7 | 6 | 5 | 6 | 6 | 6 | 4 | 6 | Logs as files with a level of their own; the data-folder move; the update chain, the cache deletion and the HDR proxy name are as they were. |
+| concat (Rust) | 7 | 4 ↓ | 5 | 6 | 6 | 6 ↓ | – | 5 | `studio.rs` 10 029; ten of twelve worker paths epoch-gated, art and save not; Save Audio As is window-only; a full publish per frame. |
+| concat (Slint) | 7 | 6 | 6 | 7 | 7 | 6 ↓ | – | – | Seeded-not-bound held everywhere; 0 ⌘ glyphs now; still 2 `accessible-*` properties, 6 literal strings, 28 colours outside `theme/`, no keyboard path into any dialog. |
+| concat-api | 8 | 8 | 5 | – | 8 | 6 ↓ | 6 | 7 | One line changed in six days while the window gained five operations the API cannot name; home-folder root and open-then-save unchanged. |
+| concat-server | 7 | 8 | 5 | – | 8 | 8 | 6 | 7 | Zero churn: fd leak, pre-auth seats, no `catch_unwind` on the Hub, gRPC without the JSON caps. |
+| concat-cli | 7 | 8 | – | – | 7 | 6 ↓ | 6 | 3 | Installs no logger, so the week's export diagnostics never reach a `serve` user; `serve` still parks forever. |
+| concat-speech | 7 | 5 | 6 | 7 | 6 ↑ | 7 | 7 | 6 | One `accelerated()` switch and one fallback serve CoreML and DirectML alike; nothing logs which provider a session got; `tts.rs` doc says whisper is a child process. |
+| concat-vision | 8 | 7 | 5 | 6 | 9 | 8 | – | 7 | Zero churn since the 23rd. |
+| concat-text | 8 | 7 ↓ | 5 | 6 ↑ | 9 | 8 | – | 7 | Faces mmapped once behind `Arc` (28 Sept #20 closed); shapes appended to the one 1 974-line file; emoji paint as boxes on a Mac. |
+| concat-android | 8 | 7 | – | – | 7 | 7 | – | 1 | Portrait locked; 10 `unsafe` each argued; still no tests; ships Chatterbox against the speech crate's own words. |
+
+---
+
+## 4. Scores per feature
+
+Status: **done**, **partial** (works with a known gap), **thin** (present,
+weak). New rows since the 28th are marked ★.
+
+| Feature | Where | Status | 28 Sept | Today | Why |
+|---|---|---|:-:|:-:|---|
+| Launcher, new project, recents | `ui/start.slint` | done | 8 | 8 | Defaults to the desktop the system names; chip rows have arrow keys and a focus ring. |
+| Import, bin, drag and drop | `ui/workspace/media-pane.slint` | done | 8 | 8 | Drop works on native Wayland; "N selected" and the empty-bin sentence are untranslated. |
+| Timeline editing | `ui/timeline/`, `panes/timeline.rs` | done | 8 | 8 | ★ Razor B / Ctrl+B snaps to the frame and refuses within `MIN_DURATION`; ★ scrollbar whose extent never shrinks under a drag. |
+| ★ Volume line | `lanes.slint:587-675`, `studio.rs:6205` | partial | – | 6 | One drag is one undo step; it does not flush a pending title commit first (finding 9); `Colors.white` not a token. |
+| Undo and redo | `concat-project/src/editor.rs` | done | 8 | 8 | Sharing survives split, ripple and edit (two `ptr_eq` tests now); every snapshot still clones the bin. |
+| Transitions | `concat-render/src/plan.rs`, `concat-export/src/lib.rs:746` | done | 8 | 7 ↓ | ★ A dissolve into a clip with no handle holds its first frame; but the hold runs the clip's keyed effects early, and keys on a clip with a handle play early by the transition's length (finding 7). "Keep alpha" is unobservable. |
+| Effects and filters | `concat-effects/packages/` | done | 8 | 8 | 98 GPU packages, 344 probes, shapes round and glows past white; `[[replaces]]` drops keys and is now a hijack vector (finding 2). |
+| Colour and HDR | `model.rs`, `gpu.rs`, `encode.rs` | done | 7 | 7 | HLG and PQ round trips pinned; hardware HDR encodes carry static metadata in the container only; pre-0.2.5 imports still never re-probed. |
+| Scopes | `concat-render/src/scopes.rs` | done | 7 | 7 | Verified this time: constants agree between shader and drawer, pinned by test; one failed readback freezes them for good (low). |
+| Colour grading | `packages/concat.color-wheels`, `grading.slint` | done | 7 | 7 | Unchanged: wheels keyable, curves not, pointer-only. |
+| Titles and text | `concat-text`, `concat-host/src/titles.rs` | done | 7 | 7 | ★ Wraps at the frame edge with no box, circular plate corners, ★ several titles styled together; emoji are boxes on macOS (finding 14); missing font silent. |
+| ★ Shapes / Stickers | `media-pane.slint:1415`, `clips.rs:122`, `concat-text/src/lib.rs:1173` | partial | – | 6 | Seven figures placed in one command, round-trip and tidy; cannot be edited after placement, deleted by 0.2.5 on save, unknown kinds become squares (finding 5). |
+| Keyframes, animations | `concat-project/src/model.rs` | done | 7 | 7 | `SetClipKey` no longer records a no-op; a drag on a keyed position keeps its key; preset animations of 0.2.4 still dropped silently. |
+| Crop, transform, blend, masks | `concat-render/src/plan.rs` | done | 7 | 7 | Darken and Lighten parity at 0.99; Screen and Add never held to the oracle. |
+| Cutout, smart brush | `concat-vision/src/segment.rs` | done | 7 | 7 | Unchanged. |
+| Enhance | `concat-host/src/enhance.rs` | thin | 6 | 6 | Copies still land in `cache/`. |
+| Auto-captions | `ui/dialogs/captions.slint` | done | 7 | 7 | Epoch-gated end to end; the transcript's words sit off by the stream start on MTS files (finding 3). |
+| Text-to-speech | `concat-speech/src/tts.rs` | done | 7 | 7 | ★ Chatterbox on DirectML with a CPU fallback; nothing says which provider ran; `DirectML.dll` is not staged beside the binary. |
+| Templates | `concat-host/src/templates.rs` | done | 7 | 7 | Poster copied under a name nothing reads (low). |
+| LUT import | `concat-effects/src/cube.rs` | done | 8 | 8 | Unchanged. |
+| Custom and community packages | `concat-effects/src/catalogue.rs` | done | 7 | 6 ↓ | Trial at 512 px / 3 s exists, at the knobs' defaults only, on the monitor's own compositor; `[[replaces]]` unfenced (findings 2, 6). |
+| Proxies | `concat-host/src/proxy.rs` | partial | 5 | 5 | ★ Full quality plays the files themselves; a pre-0.2.5 HDR import with no proxy gets an HDR copy under its SDR name. |
+| Export | `ui/dialogs/export.slint`, `concat-export` | done | 7 | 7 | ★ Survives a dead device on a fresh one, then the software adapter; ★ resolution is the timeline's own; ★ picture and scrub strip in the sheet (coalesced, not per pixel); rate ladder is a fixed 24/30/60 (finding 8); still one thread. |
+| ★ Hardware encode | `concat-media/src/encode.rs:325` | done | – | 6 | NVENC, Quick Sync and AMF lead for HEVC and AV1; fallback at open only; nothing on this machine or in CI exercises a successful chip open; "· hardware" claimed for Media Foundation's software transforms. |
+| Playback and audio | `concat-host/src/playback.rs`, `concat-media/src/samples.rs` | partial | 6 | 5 ↓ | The preview's sound runs ahead of the picture by the stream start on MPEG-TS files (finding 3); the mixer still scans every clip per sample. |
+| Hardware decode | `concat-media/src/hardware.rs` | done | 7 | 7 | On by default everywhere with a device, not only Windows and Android as the 28th said; the "5× slower" row measures a path the app never takes for 8-bit H.264. |
+| Remote API | `concat-api`, `concat-server` | partial | 6 | 6 | One line changed; the window's new operations are five reachable, four not; `export.run` cannot refuse a hardware encoder. |
+| Self-update | `concat-host/src/updates.rs` | done | 5 | 5 | ★ `.msi` and Store detection; same trust chain, downgrade still one click; any executable under Program Files is taken for an `.msi` install. |
+| ★ Logs | `concat-host/src/logs.rs`, `studio.rs:2136` | done | – | 7 | A file per run, ten kept, 16 MiB cap, Concat's own level separate from the crates underneath; bursts folded in the window; a burst followed by quitting is never counted. |
+| ★ Keyboard menus, platform keys | `app.slint:418`, `platform.rs:473` | partial | – | 6 | F10 / Alt+letter open and walk; arrow walking stalls past four dead rows; Exit on Windows, Quit on Linux. |
+| ★ Save Video Frame As, Save Audio As | `studio.rs:6105`, `5488` | done | – | 7 | Both on a worker, epoch-gated, dialog on the next loop turn; the extension is substituted after the overwrite prompt; Save Audio As is window-only. |
+| Phone shell | `ui/phone/`, `concat-android` | partial | 6 | 6 | ★ Sheets close on a swipe; portrait locked; the shell's own 44 px rule broken by its chips and transport buttons; no preview on the phone's export sheet. |
+| Relink missing media | `ui/dialogs/relink.slint` | done | 8 | 8 | |
+| Themes and accent colours | `ui/theme/` | done | 8 | 8 | ★ Lime is the default; light and dark token sets identical. |
+| Languages | `concat/locales/` | partial | 8 | 6 ↓ | 1 001 keys × 14 locales, 0 missing; but 128 to 197 values per Latin and CJK locale are the English string verbatim and the check cannot see it (finding 4); no RTL. |
+| Accessibility | the whole UI | thin | 2 | 2 | Still two `accessible-*` properties in 28 796 lines; 149 `TouchArea`s, 14 `FocusScope`s. |
+| ★ Windows data folders | `concat-host/src/dirs.rs:95` | done | – | 6 | Models and caches in `%LOCALAPPDATA%`, moved once by rename; a failed move is silent and orphans gigabytes in the roaming profile. |
+
+---
+
+## 5. The 28 September findings, re-verified
+
+Of the 18, none is closed outright, four are partly closed, and fourteen
+are open. Two of the open ones are worse than described (7 and 8) and one
+is sharper (11). The engine crates the findings live in barely changed in
+six days; the work went into the window.
+
+| # | Finding (28 Sept) | Status | Evidence today |
+|:-:|---|:-:|---|
+| 1 | Self-update trusts a digest beside the package; no signature; quarantine stripped; Windows unsigned | **open** | Releases from `api.github.com` (`concat-host/src/updates.rs:45`), package and digest from the same release's `manifest.json` (`updates.rs:440-456`), `xattr -dr com.apple.quarantine` (`596-598`). `a387328` signs the Windows installer only when `WINDOWS_CERTIFICATE_PFX` is set (`build-app.yml:666-684`, "unsigned otherwise"); `0d5bda5` adds the `msiexec /i` path (`updates.rs:533-539`). |
+| 2 | Remote page rooted at the home folder; no `HOME` = unconfined; open-then-save | **open** | `concat-server/src/lib.rs:145-149`, `80-86`; `concat-api/src/lib.rs:244` returns `Ok(())` on empty roots; Remote page passes no roots (`concat/src/panes/settings.rs:573-577`); `project.save` unchecked (`lib.rs:454-456`). Byte-identical since the 28th. |
+| 3 | Every JSON connection leaks an fd; seats counted before auth | **open** | `concat-server/src/json.rs:102` `Seat::take` before any read; `106-122` closer pushed into `connections`, drained only in `shut` (`lib.rs:284-288`); `json.rs:100` spins on EMFILE. |
+| 4 | Clear cache deletes the media behind reversed and enhanced clips; never swept | **open** | `reverse.rs:88-93`, `enhance.rs:71-74` name copies under `project/cache/`; `projects::clear_cache` removes the folder (`projects.rs:330-338`) from `studio.rs:6703`; `proxy::sweep` touches `cache/proxy/*.mp4` only. |
+| 5 | 0.2.4 documents with reversed clips or preset animations open silently changed | **open** | `MIGRATIONS` still two steps (`concat-project/src/doc.rs:62-71`); nothing reads `"reverse"` or `animation*` but a menu arm (`studio.rs:9627`); CHANGELOG 0.2.5 "Changes to note" does not say it. |
+| 6 | Embedded API opens a second GPU device | **open** | `settings.rs:581-586` builds `Api::new(events)` and shares only the exporter and the register; `export.run` → `run_on(request, None)` → `best_compositor()` → `WgpuCompositor::new()` (`concat-export/src/lib.rs:1240-1243`, `1165`). The new `recovered()` path does cover API exports. |
+| 7 | `[[replaces]]` drops keys, is silent, skips user packages and templates | **open, worse** | `Package::replacing` returns `keys: BTreeMap::new()` (`concat-effects/src/catalogue.rs:591`); `session.rs:105` discards the bool; no test. The 28th's "a user package's `[[replaces]]` never fires" is wrong today: `Catalogue::builtin()` returns `CURRENT`, which `install_with` fills with user packages (`catalogue.rs:901-903, 949-950`), so it fires at every open, and nothing checks what it replaces (finding 2). |
+| 8 | HDR proxies not regenerated for pre-0.2.5 imports | **open, worse** | No probe at open (`studio.rs:6608-6620` keys by the document's `color_space`); `proxy::write` decides HDR from the probe (`proxy.rs:208`) while the name comes from the document (`proxy.rs:114`), so an HDR import from 0.2.4 with no proxy yet gets an HEVC/HLG copy under its SDR name. |
+| 9 | Hub: heavy work in line, no `catch_unwind`, slow subscriber dropped silently | **open** | `concat-server/src/hub.rs:83-87`; probes and composites on the API thread (`concat-api/src/lib.rs:326, 465, 485, 632, 690-704`); outbox `sync_channel(256)` and the silent drop (`json.rs:304, 319-325`, `hub.rs:58-66`). |
+| 10 | Downgrade is one click with no gate | **open** | `install_version` takes `releases.get(version_choice)` with no comparison against `Version::current()` (`settings.rs:461-473`); the list keeps every release ≥ 0.2.5 (`updates.rs:141-143`). |
+| 11 | No ceiling on frame size or rate in the model | **open, sharper** | `is_sane` is `> 0` (`model.rs:1917-1919`); `rate_num` is `i64` and the export's `FrameRate::new(Rational::new(rate_num, rate_den))` goes through `reduce`, which is `.expect("rational overflowed i64")` (`concat-core/src/time.rs:100`, `234`). A hand-edited `"rateNum": 9000000000000000000` loads and the first `duration × rate` panics the export thread. |
+| 12 | Every undo snapshot deep-copies the bin | **open** | `editor.rs:141` clones the project; `media: Vec<MediaItem>` (`model.rs:1941`); `Batch` clones again per level (`commands/mod.rs:1027`). Batch title styling now sends one `Batch` of N `UpdateClip`s per gesture, so the pattern is exercised more. |
+| 13 | Export one thread with a readback per frame; 400-piece curves; mixer scans every clip | partial | Loop unchanged (`concat-export/src/lib.rs:1301-1468`); `ExportClip` 41 fields; the window no longer owns the audio cut (`studio.rs:2339, 5538` call `concat_export::audio_pieces`), the `clamp(8, 400)` lives in one place now (`lib.rs:1077`); `playback.rs:1032` unchanged. |
+| 14 | Font by absolute path; missing font silent | **open** | `fonts.add_file(...)` result dropped (`titles.rs:339-342`); `pick` returns whatever answers (`concat-text/src/lib.rs:332-339`); `Rendered` carries no "face used". |
+| 15 | Small correctness in concat-project | partial, grown | (b) closed: `SetClipKey` compares before and after (`properties.rs:169-173`). Open: `has_non_finite` lists 8 of 11 `TextStyle` f64 fields (`commands/mod.rs:912-920`); `follow_first_hdr` re-flip (`editor.rs:238-244`); 5 of 20 structs carry `extra`, `ShapeStyle` joins the ones without (`model.rs:901`); `"concat": "0.1.0"` (`doc.rs:320`). |
+| 16 | Server hygiene: socket umask; `connections()` over-counts; `stage` names; `project.create` unbounded | **open** | `lib.rs:195-197` no `set_permissions`; `lib.rs:257-259`; `message.rs:591` vs `concat-export:979, 984, 1466`; `lib.rs:385-394`. |
+| 17 | Host hygiene: `rpm -q` on the UI thread; predictable temporaries; two TLS stacks | partial | `installed_as` memoised in a `OnceLock` (`updates.rs:253-255`), once per process: closed. Temporaries unchanged (`concat-export/src/lib.rs:957-958`); `Cargo.lock` still has ureq 2.12.1 on rustls and ureq 3.4.1 on native-tls, `webpki-roots` twice. |
+| 18 | Carried: 5 ⌘ glyphs; `ui/demo/`; dead `core::Project`; `Path::exists` in the wasm crate; speed range ×5; `mobile.yml` sentence | partial | ⌘ **closed**: 0 in strings, labels from `@keys` and `platform::keys` (`tray.slint:138`, `platform.rs:473-477`). `mobile.yml` **closed**: the sentence is true, `mobile.yml:36` has `pull_request:`. `ui/demo/` partial: `editor-chrome.slint` no longer imported by the live tree; `effects-panel.slint` still is, for one 3-field struct (`app.slint:21`, `editor.slint:4`, `config-pane.slint:9`), 4 files / 738 lines. Open: `core::Project` (`concat-core/src/timeline.rs:499-514`); `Path::exists` (`model.rs:2028-2031`); speed range 3 definitions + 3 literals (`studio.rs:4305, 8573`, `cutout.rs:178`). |
+| – | Two window tests flaky by locale | **open** | `format.rs:496-504` and `panes/speech.rs:668-679` read `t()` with no lock; the i18n tests select `de` under a lock private to their module (`i18n.rs:377-380, 516, 549`). CI runs the window binary with default test threads (`ci.yml:112` is build jobs), so it can flake there too. |
+
+---
+
+## 6. New findings, by severity
+
+No new critical. Each carries a location and whether the reviewer read the
+whole path **(verified)** or reasoned from part of it.
+
+### High
+
+1. **`main` has been red at `cargo fmt --check` for six days, and nothing
+ after it has run.** (verified) Every completed CI run since `a262553`
+ (28 September 22:16 UTC) fails at `ci.yml:93`; the gates behind it,
+ clippy under `-D warnings`, `locales.py --check`, `models.py --check`,
+ the whole test suite, the quick perf budgets and the gRPC build
+ (`ci.yml:98-126`), have not executed on `main` for 71 commits. The first
+ unformatted sites came in the afternoon of the 28th
+ (`concat-text/src/lib.rs:436, 717` from `35b80d1`,
+ `concat-media/src/encode.rs:454, 1531, 1546` from `8d3bf19`,
+ `concat-host/src/dirs.rs:95` from `dc12b64`) and sixteen more followed
+ (`f97c745`, `0699b44`, `21855a7`). The gate is on push to `main`
+ (`ci.yml:5-7`), it fired on every push, and it was ignored. A 0.2.6 tag
+ today runs `release.yml:66-68`, fails at fmt and produces no bundles;
+ meanwhile a test regression in any of the 60 commits is invisible. A
+ related gap: the engine job installs `stable` (rustc 1.99.0 in the log
+ of run `37165478921`) and never reads `src/rust-toolchain.toml`'s 1.93
+ pin, so CI and a local checkout format and lint with different
+ toolchains. The fix is `cargo fmt` and one commit; the lesson is that a
+ gate nobody reads is not a gate.
+
+2. **A user package can hijack every link to a live built-in.** (verified)
+ `Replaces::names` matches any id (`concat-effects/src/manifest.rs:91-93`);
+ neither `Manifest::validate` (`manifest.rs:489-504`) nor
+ `Catalogue::add` (`catalogue.rs:986-1003`) checks that the replaced id
+ is absent from the catalogue; `Catalogue::upgrade` takes the first
+ package whose table matches (`catalogue.rs:1050-1058`);
+ `Catalogue::builtin()` returns the installed catalogue once
+ `install_with` has run (`catalogue.rs:901-903, 949-950`); and
+ `Session::open` applies the upgrade before any edit, outside undo, with
+ no notice (`concat-host/src/session.rs:104-105`,
+ `concat-project/src/editor.rs:89-107`). A shared package `evil.glow`
+ declaring `[[replaces]] id = "concat.glow"` is installed; every project
+ opened afterwards has its Glow links rewritten to `evil.glow` with their
+ keys dropped, and the next save writes that out. The trial at install
+ vets the shader, not the manifest's claims. Fix: refuse a `[[replaces]]`
+ whose id or bare id the catalogue already answers to, at `add` and
+ `install_with`, and tell the person when `upgrade_links` changed
+ anything.
+
+3. **The preview's sound and the transcript ignore the stream start; the
+ export does not.** (verified) `concat-media/src/samples.rs:176-182`
+ seeks to `start × AV_TIME_BASE` with no `start_of`, and the graph trims
+ on raw timestamps (`atrim=start={start}`, `samples.rs:91, 95`). For an
+ MPEG-TS or MTS file whose first sample is at 1.4 s, `start = 0.5` lands
+ at the file head and `atrim` keeps every frame, so the cached PCM begins
+ at the sound's first sample while the picture begins 0.5 s in: the
+ preview's sound runs ahead of the picture by `min(start, 1.4)` s
+ (`concat-host/src/playback.rs:734-745`), and the transcript's words sit
+ the same amount off (`concat-speech/src/transcribe.rs:339-350`). The
+ export mix adds the start back on the seek and measures its lead off the
+ first frame (`concat-media/src/audio.rs:422-434`), so the file is right
+ and the preview is wrong: two definitions of source time, which
+ `audio.rs:11-14` says do not exist. The 28th's "stream start time
+ ignored" was closed for video only. `fixtures/starts-at-1.4s.ts` has no
+ audio stream, so the test needs a new fixture.
+
+4. **Eight locales are a fifth English, and the check cannot see it.**
+ (verified) `locales.py --check` compares key sets
+ (`scripts/locales.py:134-148`), so a value equal to the English string
+ passes. Today, ignoring symbol-only strings: de 195, fr 197, pt-BR 170,
+ es 168, it 168, tr 150, ja 137, ko 128 values equal to `en.json`; hr 53,
+ fa 5, zh-TW 5, zh-Hans 4, ru 3. In `de.json`, `captions.addCaptions` is
+ "Add captions", `captions.oneTitlePerSentence` the full English
+ sentence, and every `effects.*.description`. A German user opens the
+ Captions sheet and reads English. The 33 keys added since the 28th are
+ all translated; the debt is older (193 in `de` at `73e2176`). A
+ `--check` that flags a value equal to English on a key of two or more
+ words would make the "all 100 %" of the README mean what it says.
+
+### Medium
+
+5. **Shape clips do not survive 0.2.5, and an unknown shape kind is
+ rewritten as a square.** (verified) `wire::clip_kind` reads `"shape"`
+ as `Video` (`concat-project/src/model.rs:1159-1162`), `settle_clip` drops
+ a Video clip with no media (`doc.rs:266-269`), and the flat `clips`
+ mirror carries them too (`doc.rs:345-348`): 0.2.5 opens a 0.2.6 project
+ with every shape gone and the next save removes them, at the same
+ `DOCUMENT_VERSION` of 1 (`doc.rs:41`), with no "newer" refusal. In the
+ other direction `wire::shape_kind` falls back to `ShapeKind::default()`
+ (`model.rs:1184-1187`) and serialises it as `"square"`, so a later
+ build's `"star"` is a square here and `"star"` is gone on save.
+ `TextAlign` and `ColorSpace` have the same lossy fallback. Also: no
+ command edits `Clip::shape` after placement (`ClipPatch` has `text`, no
+ `shape`, `commands/mod.rs:143`), so fill, stroke and figure are fixed
+ at `AddShapeClip` and an API caller has delete-and-re-add; and
+ `AddShapeClip` is one of 17 of 44 commands absent from the serde
+ round-trip test (`lib.rs:3440-3565`).
+
+6. **The install trial runs at the knobs' defaults only, and a timed-out
+ trial kills the monitor.** (verified) `trial_pass` is
+ `pass(&BTreeMap::new(), None)` (`concat-effects/src/catalogue.rs:489-491`)
+ and `bounded` asks only that a `break` exist in the loop
+ (`shader.rs:1108`): a loop whose count is `params.n × 1e6` with default
+ 1 passes the trial and hangs the device when the slider is raised. And
+ `trial_at` on timeout sets `self.dead = true` (`concat-render/src/gpu.rs:1869`)
+ on the monitor's own compositor (`concat-host/src/preview.rs:120-126`);
+ from then every frame answers "the GPU device was lost"
+ (`preview.rs:234-237`), the pool stays allocated, and nothing in
+ `concat/src` rebuilds the monitor. A hostile package is refused and
+ still ends the session's preview. `Package::params_at(At::Max)` exists
+ (`catalogue.rs:597`) and the vet does not use it; `monitor.sibling()`
+ exists and the trial does not run on one.
+
+7. **The dissolve hold plays the clip's keyed effects early, and keys on a
+ clip with a handle play early by the transition's length.** (verified
+ for the hold; partial for the handle case) `pre_roll` clones the
+ incoming clip (`concat-export/src/lib.rs:776`), clears its placement
+ keys (`803`) but keeps `effects` with their knob keys, and `passes_at`
+ resolves them at the hold's own fraction (`lib.rs:1478-1495`): a film
+ grain keyed 0→100 over a 4 s clip, cross-faded in over 1 s with no
+ handle, runs its ramp inside the one-second hold and snaps back to 0 at
+ the cut. With a handle, `pre_roll` extends the clip backwards
+ (`lib.rs:762-767`) and keys are fractions of the clip, so a key at
+ fraction *a* lands *d·(1−a)* earlier than the clip's own span says.
+ Monitor and export agree with each other, not with the keyframe markers.
+ The holds are unit-tested (`lib.rs:2444-2512`) but not pixel-pinned;
+ `concat-host/tests/export.rs` has not changed since the 28th.
+
+8. **The export's frame rate is a fixed 24/30/60 ladder, not the
+ timeline's.** (verified) `EXPORT_RATES` (`concat/src/studio.rs:154`),
+ `rate: 1` default (`panes/export.rs:113`), options hard-coded
+ (`ui/dialogs/export.slint:363`). `e4d3507` made the resolution the
+ timeline's own and greys the other rungs (`export.slint:350-356`); the
+ rate was left. A 25 fps PAL or 23.976 cut exports at 30 fps unless the
+ person notices, with a frame conversion the sheet never names.
+
+9. **A volume-line drag can drop typed title words.** (partial: the UI half
+ read) `clip_set_text` leaves the words on the echo with
+ `commit_pending = true` and `commit_targets = [title]`
+ (`studio.rs:4446-4448`); the volume `TouchArea`'s press does not go
+ through `clip_pressed`, so nothing flushes; `clip_volume` then
+ overwrites `commit_targets` with the sound clip (`6213-6215`), and when
+ the 120 ms timer fires `commit_now` diffs only that clip
+ (`4634-4644`). Type into a caption's Content field, then press and drag
+ a sound clip's volume line within 120 ms of the blur: the words go with
+ the echo. Every other gesture entry flushes first (`3926`, `4883`,
+ `4967`, `lib.rs:927`).
+
+10. **Art caches are never released across projects, and the art workers
+ are not epoch-gated.** (verified) `open_project` and `close_project`
+ clear `art_failed`, `art_stamp` and `proxied` only
+ (`studio.rs:6626-6628`, `6739-6741`); `peaks`, `strips`, `windows`,
+ `media.thumbs` and `waves` live for the process. `spawn_art` and
+ `spawn_strip` go through `deliver(None, …)` (`host.rs:253-275`;
+ `studio.rs:2822`, `2925`), so a strip decoded for project A lands in
+ B's tables after the switch; `save` uses the ungated `spawn`
+ (`studio.rs:2310-2320`), so a failed save of A sets `dirty` on B;
+ `close_project` cancels cutouts, enhancers and reversers but never a
+ running export (`6712-6762`). Ten of twelve worker paths are gated; the
+ 28th's finding 3 is closed for those and open for these.
+
+11. **A full publish per monitor frame.** (verified) Every `deliver` and
+ `on_ui` ends in `publish` (`host.rs:243`, `303`) and monitor frames
+ arrive through `spawn_in_project` (`panes/monitor.rs:257`), so playback
+ is one full publish per frame, about 30 a second, on top of the
+ transport timer's `publish_lanes` every 33 ms (`studio.rs:2530-2566`).
+ One full publish rebuilds every visible `ClipData` with wave and strip
+ lookups (`7199-7248`), the key editor rows (`7325`), the font list
+ collected and sorted (`4456-4480`), the bin rows (`8785`), the export
+ ladder three times (`panes/export.rs:470-499`), the settings data and
+ both menus. `refresh_art` itself is one comparison now
+ (`8910-8920`); the publish behind it is not.
+
+12. **No dialog button, card or sheet is keyboard-reachable, and Escape
+ closes nothing.** (verified) `Button` (`ui/primitives/modal.slint:168-251`)
+ has a `TouchArea` and no `FocusScope`; `QualityCard`, `PillButton`,
+ `SheetButton`, `ToolItem`, every `MediaCard`, `CatalogueCard` and
+ `ShapeCard` likewise. `Key.Escape` is handled only in `select.slint:308`,
+ `context-menu.slint:192`, `number-field.slint:143` and
+ `tabs.slint:164`; `Modal` (`modal.slint:17-160`) and `ExportDialog`
+ have no key handling. Open Export with Ctrl+E: Tab reaches the Select
+ rows and nothing else, Escape does nothing, Return cannot press
+ Export. Together with the two `accessible-*` properties this is the
+ accessibility row's 2.
+
+13. **Six user-visible strings and the whole effects category bar bypass
+ `I18n`.** (verified) `tray.slint:182` "Merge selected",
+ `media-pane.slint:2197` "{n} selected", `media-pane.slint:2239` the
+ empty bin's only sentence, `media-pane.slint:576` "still",
+ `text-panel.slint:179` "Empty", `controls.slint:100` "% of height";
+ and `CategoryPillBar`'s "Featured", "Retro & Film", "Optical & Lens"…
+ (`media-pane.slint:1332-1340`), English on every locale and compared by
+ string to Rust (`studio.rs:312`). `locales.py` extracts `I18n.t(...)`
+ calls only, so none of these can reach a translator.
+
+14. **Emoji paint as boxes in titles on macOS.** (verified) `draws` counts
+ a face only when the glyph has an outline
+ (`concat-text/src/lib.rs:530-545`); Apple Color Emoji has none, so no
+ face covers "🎉"; `assign` falls through to index 0 (`lib.rs:432-452`)
+ and the style's face paints its `.notdef` box, in the monitor and in
+ the export. Segoe UI Emoji has outlines, so Windows is fine; Linux
+ depends on the fonts installed.
+
+15. **No Unreleased section, and the version is still 0.2.5.** (verified)
+ `CHANGELOG.md` has one `##` section; nothing touched it in 60 commits;
+ `src/Cargo.toml:6` says 0.2.5. Not recorded: hardware encode, Chatterbox
+ on DirectML, **models and caches moved to `%LOCALAPPDATA%`** (a
+ migration on first launch), keyboard menus and Exit/Quit, the razor,
+ the scrollbar, the volume line, Save Video Frame As, the export sheet,
+ the activity-trail log, multi-title styling, Stickers, the lime accent,
+ Wayland drop, export surviving a lost device, the dissolve hold, the
+ Store `.msix`, the AppImage rename. The release script's fallback is the
+ commit list, which the project's own rule says is not wanted.
+
+16. **The code grew without tests.** (verified by `git log --numstat`)
+ Since the 28th: `concat` +2 111 Rust and +1 291 Slint lines, 5 new
+ tests (`platform.rs` 2, `wayland_drop.rs` 2, `studio.rs` 1);
+ `concat-media` +126, 0; `concat-speech` +18, 0; twelve packages' WGSL
+ changed, covered by probes only; the logs rewrite has one test
+ (`level_names_round_trip`). 26 tests workspace-wide for 3 800 lines,
+ none of them end to end. Nothing covers `commit_now`/`flush_commit`
+ ordering, `clip_volume`, `save_frame_as`, `split_clip_at`, a real
+ device loss, or a successful NVENC/QSV/AMF open.
+
+17. **The README sells an MCP transport that does not exist.** (verified)
+ `README.md:62` "a JSON-RPC, gRPC and MCP API", `:79`, `:114`. No crate,
+ feature or binary serves MCP; the only hit is a doc comment
+ (`concat-api/src/lib.rs:7`); `version.capabilities` can name `json-rpc`,
+ `unix-socket` and `grpc` only (`concat-server/src/lib.rs:137-141`).
+ An agent builder reading the README looks for an endpoint that is not
+ there.
+
+18. **The CLI installs no logger, so the week's export diagnostics are
+ dropped.** (verified) `concat-cli/Cargo.toml:14-27` depends on neither
+ `log` nor `concat-host::logs`; `main.rs` has no `set_logger`. `0699b44`
+ and `d04a34c` put the device-loss reason, the adapter name and the
+ recovery on `log::warn!`/`log::info!`
+ (`concat-export/src/lib.rs:1206-1211, 1247-1252`); through `serve` or
+ `api` they go to the no-op logger and the caller sees only
+ `export.failed` with the last string. The Settings log level has no
+ CLI or API counterpart either.
+
+19. **The API fell six days behind the window.** (verified) One line
+ changed in `concat-api` (`proxy: false` on `FrameSpec`, `lib.rs:700`),
+ none in `concat-server` or `concat-cli`, while `message.rs:25-26` still
+ promises "every operation the window can perform is one an API caller
+ can". Reachable: split (`SplitClips`), clip level (`UpdateClip.volume`),
+ shapes (`AddShapeClip`, though `name` is a required string the caller
+ must invent), Save Frame (`preview.frame`), the resolution ladder. Not
+ reachable: Save Audio As, the log level, the activity trail, CBR or a
+ bitrate (`lib.rs:568-569` hard-codes VBR), and the hardware encoder,
+ which `export.run` cannot refuse because `EncodeOptions` gets
+ `hardware: true` unconditionally (`concat-export/src/lib.rs:1277`) and
+ neither `ExportRequest` nor `ExportSpec` carries the field; the window
+ only labels it (`panes/export.rs:545-546`).
+
+### Low
+
+20. **concat-project.** `has_non_finite` runs after the snapshot clone
+ (`editor.rs:141` vs `commands/mod.rs:1021-1023`), so a refused NaN
+ still paid for a project clone; `AddShapeClip` repeats `AddTextClip`'s
+ lane-minting block verbatim (`commands/clips.rs:131-150` = `81-100`);
+ `core::Clip::source_duration` ignores `hold` and has no caller
+ (`concat-core/src/timeline.rs:314-320`).
+
+21. **concat-render and concat-effects.** A scope readback that fails
+ freezes the scopes for good (`gpu.rs:1777-1781`, `1623-1629`,
+ `1790-1802`); `transitions.rs` is a 601-line `#[cfg(test)]` oracle
+ nothing compares the GPU against (`lib.rs:28`, `reference.rs:41-43`);
+ the built-in pipelines in `with_device` and the scopes compute pipeline
+ are built outside any error scope (`gpu.rs:977, 1014, 1059, 1122,
+ 1683-1692`), so ARCHITECTURE's "never an uncaptured error" holds for
+ package pipelines only; `read_back` polls with no timeout
+ (`gpu.rs:3441`); "transitions keep alpha" cannot be observed because
+ both stages clear opaque black and `combine` reads back opaque frames
+ (`gpu.rs:2932, 2891, 3483-3485`); the window now takes
+ `adapter.limits()` (`concat/src/gpu.rs:94-98`) and concat-render never
+ checks a plan against `max_texture_dimension_2d`; `matches_category`
+ hard-codes nine package ids (`catalogue.rs:437-475`).
+
+22. **concat-media.** No fallback once a hardware encoder has opened: a
+ `send_frame` error ends the export with no trailer
+ (`encode.rs:1235-1237`, `concat-export/src/lib.rs:1457-1459`);
+ "· hardware" is claimed for Media Foundation, whose `h264_mf` opens on
+ any Windows machine through software transforms (`encode.rs:189-193,
+ 371-373`); hardware HDR encodes carry the mastering display in the
+ container atoms only, no SEI (`encode.rs:522-541, 667-696`); the
+ colour round-trip tests fill frames with greys, which any matrix
+ survives, and `sws_setColorspaceDetails`'s return is ignored
+ (`encode.rs:1119-1128`, tests `1687`, `1830-1884`); a fresh reader
+ seeks twice (`pool.rs:367`, `918-921`); AMF is handed `qp_b` on HEVC
+ and AV1 (`encode.rs:656-662`), silently dropped; ARCHITECTURE.md:260
+ describes a hardware-decode Settings switch that `settings.rs:206-211`
+ says is "Not a choice".
+
+23. **concat-host and concat-export.** A failed data-folder move is
+ silent and leaves gigabytes in the roaming profile for good
+ (`dirs.rs:119-122`); any executable under Program Files without an
+ uninstaller is taken for an `.msi` install and `msiexec /i` installs a
+ second copy (`updates.rs:287-309`); a template's poster is copied as
+ `cache/preview.jpg` while `poster_cache` reads `poster.jpg` and deletes
+ `preview.jpg` (`templates.rs:349-353`, `media.rs:651-653, 705`); a
+ burst's count line is written only when the next message arrives
+ (`studio.rs:2136-2169`), so a burst followed by quitting is never
+ counted; `dirs.rs:149-154` is a unit test that on Windows runs
+ `move_data` against the real `%APPDATA%`; ARCHITECTURE.md:240 says
+ proxies are "quarter-size H.264" while an HDR source's is ten-bit HEVC
+ (`proxy.rs:223-247`); `updates.rs:777`'s fixture still names a
+ `-linux-` AppImage after `c75b09c`.
+
+24. **The window.** `save_frame_as` and `export_clip_audio` substitute the
+ extension after the OS's overwrite prompt (`studio.rs:6148-6155`,
+ `5577-5584`) and on a phone write unconditionally; inspector ranges
+ live in four literal copies (`studio.rs:546, 4265, 5199, 7723`);
+ `freeze_at_playhead` decodes on the UI thread (`6342-6349`); the Remote
+ page mints its token from two `RandomState` hashers
+ (`settings.rs:856-864`) while `token::mint` one crate away uses
+ `getrandom`; the token's doc comment now sits above `LOG_LEVELS`
+ (`settings.rs:843-848`); `looks_dir`'s doc says it packs a template
+ (`studio.rs:6858-6861`); `export_clip_audio` carries three superseded
+ doc blocks (`5488-5509`).
+
+25. **The UI.** Keyboard menu walking looks at most four rows ahead
+ (`context-menu.slint:87-93`), so a fresh project's Edit menu (undo,
+ redo, rule, split, delete all dead) cannot be walked; 28 colours named
+ outside `theme/` (the transition seam chrome at `lanes.slint:766-946`
+ is palette-blind on the light theme; `Colors.white/black` at
+ `lanes.slint:615, 664, 670` are new); the phone shell's own 44 px rule
+ is broken by `LaneChip` at 34 px and the transport buttons at 40
+ (`phone/editor.slint:175-177, 455-489`), and a sheet full of 28 px rows
+ is mostly unscrollable surface (`pane.slint:124-133`); the inspector's
+ `in-chain` scans are hand-unrolled eight deep, so a ninth link draws as
+ "not applied" (`clip-inspector.slint:282-289, 708-716, 968-976`);
+ `ui/demo/` is now reachable through one 3-field struct and could go in
+ one commit; the phone's export sheet has no preview
+ (`export.slint:823-849`); `IconButton.label` and `RoundButton.label`
+ are documented as the accessible name and never bound to
+ `accessible-label` (`panel.slint:193`, `phone/controls.slint:20`).
+
+26. **The server.** gRPC has none of the JSON transport's caps: no
+ connection limit, no pre-auth timeout, nothing beyond tonic's 64-event
+ channel (`grpc.rs:70, 129-175`); `Seat` and `AUTH_TIMEOUT` are
+ JSON-only (`json.rs:48-70, 125`).
+
+27. **Speech, models and dependencies.** The phone build carries Chatterbox
+ (`concat/Cargo.toml:121` enables it for every target) against
+ `concat-speech/Cargo.toml:25-27`'s "a gigabyte of model is not for a
+ phone"; `with_execution_providers` does not fail when DirectML cannot
+ be created, so a machine without D3D12 runs on the CPU under a log line
+ saying DirectML (`chatterbox.rs:303-313`), and the Windows stage copies
+ no `DirectML.dll` (`build-app.yml:705-710`); the HF-mirror preference
+ does nothing for the seven GitHub-hosted models (`models.rs:158-163`);
+ the AVX baseline is a CI env only (`build-app.yml:85-89`), a
+ contributor's `--profile app` is AVX2 again, and no runtime CPU check
+ exists; `whisper-rs`'s `set_progress_callback_safe` leaks a box per
+ transcription (`transcribe.rs:413`); `tts.rs:7` still says whisper
+ runs as a child process. Two ONNX runtimes on every platform (pyke's
+ through `ort-sys`, k2-fsa's inside `sherpa-onnx`) and two TLS stacks
+ remain; `bzip2` is linked three times, `tiny-skia`, `fontdb`, `png`
+ and `jni` twice each.
+
+28. **CI and docs.** `WINDOWS_CERTIFICATE_PFX` and `_PASSWORD` are not
+ declared under `workflow_call.secrets` (`build-app.yml:37-53`) and
+ reach the job only through `secrets: inherit` (`release.yml:73`), and
+ signtool is given the password on the command line (`:684`);
+ `README.md:56, 80` list four platforms while `:97` adds iOS and iPadOS;
+ `src/README.md:129-130` says the quick perf set runs "on every push",
+ which has not been so since the 28th; `SECURITY.md:39-40` covers model
+ digests and says nothing about the self-update's trust root;
+ `THIRD_PARTY_NOTICES.md` names none of the eight direct dependencies
+ added since the 28th (all permissive; `unicode-script` carries the
+ Unicode data licence).
+
+---
+
+## 7. What is good, and should stay that way
+
+- **The layering, still.** FFmpeg is only in concat-media (`grep` finds no
+ other crate naming `ffmpeg-the-third`); `FrameSource`/`FrameSink` the
+ only seam; the CPU compositor a test oracle and nothing else.
+- **`recovered`: three deaths, two kinds of successor, the black frame
+ redrawn, the reason in the log** (`concat-export/src/lib.rs:1189-1213`,
+ `concat-render/src/gpu.rs:3439-3455`), and the proof that it cannot
+ write a partial output: the silent file is renamed or muxed only after
+ success and removed on error (`lib.rs:960-982`).
+- **`hold` as one boolean read in exactly the two places every renderer
+ agrees on** (`concat-core/src/timeline.rs:295-297, 325-327`), with the
+ export building the held copy rather than the model learning about
+ transitions.
+- **`settle_clip` and `Clip::tidy` extended for shapes in the right two
+ places and nowhere else** (`doc.rs:265, 275-277`; `model.rs:1511`): the
+ "one place clamps live" rule held through a new clip kind.
+- **The scheduler's `Finished` guard**: counters as a `Drop`, so a panic
+ cannot wedge the reserve, with the test that proves it
+ (`concat-media/src/prefetch.rs:374-392, 408-427`).
+- **`hardware_wins` as a measured policy with its numbers in the doc
+ comment**, and perf rows for "the app's choice" beside the raw chip
+ (`hardware.rs:193-210`, `concat-perf/src/main.rs:655-668`).
+- **Every `unsafe` block in concat-media says why it is sound** (37 sites,
+ 0 without a SAFETY line), and the new `wayland_drop.rs` is a careful
+ guest on winit's display: own queue, `prepare_read`/poll/cancel loop,
+ CLOEXEC pipe ends as `OwnedFd`, offers declined and destroyed, device
+ released on stop (`wayland_drop.rs:133-194, 338-380`).
+- **One epoch gate in one place, with a test** (`concat/src/host.rs:197-246,
+ 688-705`), and every long job's progress through `on_ui_in_project`.
+- **Dialogs from a menu row deferred to the next loop turn, with the reason
+ written down** (`studio.rs:5556-5565, 6102-6103`): the Save Audio As
+ abort was a native dialog opened inside a Slint timer callback, and
+ both Save verbs now hop through `on_ui` first.
+- **Monitor request coalescing** (`panes/monitor.rs:234-237`): one frame in
+ flight, latest wins, so the export sheet's scrub strip cannot decode per
+ pixel; and no timer runs in an idle window.
+- **Probes read back in half floats with nothing between shader and
+ number**, 344 of them across 98 packages on every `cargo test` with a GPU
+ (`gpu.rs:1885-1901`, `gpu/tests.rs:1091-1114`); the pool that is also the
+ cache, LRU within two bounds (`gpu.rs:2509-2581`).
+- **Logs as a file per run, ten kept, 16 MiB cap, Concat's own level
+ separate from the crates underneath** (`concat-host/src/logs.rs:38-41,
+ 228-268`).
+- **Faces mapped once and shared behind `Arc`**
+ (`concat-text/src/lib.rs:475-491`); per-character coverage learnt once
+ (`393-420`).
+- **The archive family's install**: `.part` → staging → `unpack_in` →
+ single rename, so nothing half-unpacked ever looks installed
+ (`concat-speech/src/tts.rs:529-580`); resumable downloads by `Range` with
+ 416 handled and a real socket test (`concat-host/src/models.rs:201-284,
+ 413-469`).
+- **`token::matches` with the test naming each promise**, used by both
+ transports (`concat-server/src/token.rs:29-36, 52-64`); `Api::writable`
+ + `resolved`; `OpenProjects` with `Drop`; `SingleFlight`.
+- **`locales.py` builds `en.json` from the source and the package
+ manifests** (`scripts/locales.py:46-96`), so a key cannot be used
+ without English and a stale key cannot survive; and the seeded-not-bound
+ discipline for every text field, with the reason at each site.
+- **The release gate order**: tag must equal the version, CI must pass
+ before a bundle is built, notes from CHANGELOG with a fallback
+ (`release.yml:46-73`); parity real on the Linux runner through lavapipe
+ and `CONCAT_REQUIRE_GPU` (`ci.yml:44-46, 116`).
+
+---
+
+## 8. Where we stand
+
+The 28 September audit ended with seven recommended steps. Six days later:
+
+| Step (28 Sept) | Landed |
+|---|---|
+| 1. Sign the update; keep quarantine; confirm downgrade; sign Windows; notarise every Mac build | Windows signing exists when a secret is set; nothing else |
+| 2. Remote page defaults, seats after auth, drop the connection clone, `catch_unwind` the Hub, lend the device | nothing (one line changed in the API, none in the server) |
+| 3. Processed media out of `cache/`; sweep art and peaks; re-probe pre-0.2.5 media | nothing |
+| 4. Say it when a document changes: migrate `reverse` and animations; surface `upgrade_links` | nothing, and a new silent seam (shapes under 0.2.5) |
+| 5. Finish the 23 Sept list | ⌘ glyphs gone; `mobile.yml` sentence was true; `rpm -q` memoised; `audio_pieces` moved to one place; the rest open |
+| 6. Pipeline the export, index the mixer, `Arc` the bin | nothing |
+| 7. Speech, enhance, proxies and CBR as API verbs | nothing; the gap grew by four window-only operations |
+
+None of the seven was taken up. What landed instead is a week of the
+window: a razor, a scrollbar, a volume line, keyboard menus on every
+platform, Save Frame and Save Audio, an export sheet with a picture, logs
+that read as what the person did, shapes, batch title styling, Wayland
+drop, Linux window edges, Windows data folders and Store packaging,
+hardware encoders, Chatterbox on DirectML, and an export that survives a
+dead GPU. Each of those is well made at the seam it touches: the new
+gestures go echo-then-command, the new workers are epoch-gated, the new
+`unsafe` is argued, the device recovery cannot write a partial file. The
+app is more pleasant to use than it was.
+
+What the week cost is of three kinds. The first is the gate: `main` has
+not been green since the release, fmt has been the only check to run for
+71 commits, and the version and CHANGELOG do not know the week happened.
+The second is the engine debt of the 28th, untouched and in three places
+worse: `[[replaces]]` went from "drops keys" to "a package can hijack a
+built-in", the HDR proxy from "not regenerated" to "written under the wrong
+name", and the frame-rate ceiling from "no ceiling" to "a panic in the
+export". The third is a new correctness seam in the two paths the audit of
+the 28th had not read in full: the preview's sound is off by the stream
+start on MPEG-TS files, and shapes vanish under 0.2.5.
+
+**Recommended order.** Each line is an hour to a few days, and each closes
+a finding above.
+
+1. `cargo fmt`, push, watch it go green; put `rust-toolchain.toml` into the
+ CI job; isolate the two locale tests; then write the Unreleased section
+ and bump to 0.2.6. (1, 15, flaky tests)
+2. Refuse a `[[replaces]]` naming a live package at install; carry keys
+ through `replacing`; tell the person when `upgrade_links` changed a
+ document; run the trial at `params_at(Max)` on a `sibling()` compositor.
+ (2, 6, 28 Sept #7)
+3. `AudioDecoder::open` takes `start_of`, seeks and trims at
+ `start + offset`, with an MTS fixture that has sound. (3)
+4. Bump `DOCUMENT_VERSION` for shapes so 0.2.5 refuses rather than deletes;
+ keep an unknown kind's name in `extra`; give `ClipPatch` a `shape`;
+ grow the round-trip test to all 44 commands. (5)
+5. A `locales.py --check` that flags values equal to English; the six
+ literals and the category bar through `I18n`; the export rate from the
+ timeline; freeze effect keys at the hold's head. (4, 8, 13, 7)
+6. Flush the pending commit in `clip_volume`; gate `spawn_art`,
+ `spawn_strip` and `save` by epoch and clear the art tables at close; a
+ `publish_lanes`-only delivery for monitor frames. (9, 10, 11)
+7. A `FocusScope` on `Button` and Escape on `Modal`; bind
+ `accessible-label` where the doc already promises it. (12)
+8. Then the 28 September list, in its order: sign the update; the Remote
+ page's root, seats and fd leak; processed media out of `cache/`; the
+ migration notices; the export pipeline; the API verbs. (28 Sept 1-7,
+ today 17-19)
+
+After those, phase 4 of the HDR plan (a true HDR preview), then phase 5
+(zero-copy hardware frames, hardware encoders on the Mac) in the plan's
+order.
+
+---
+
+## 9. Caveats
+
+Scores are judgements over code, calibrated against the 28 September
+scores but not measured against users or another editor. Every crate was
+read in full this time; nothing is provisional. Findings marked (verified)
+were read end to end by the reviewer and the three headline ones re-read
+by hand; the rest rest on one reviewer's reading and a line number you can
+open. The hardware encoders and DirectML could not be exercised on this
+machine; what is said of them is read from the code and the workflows. The
+app was not launched: "user likeability" is judged from what the UI code
+does, not from users.