Skip to content

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

docs(audit): 24 October 2026 - four area reviews, scores, and the Highs they found fixed the same day

Jareer971ad1d9214e
diff --git a/audit/2026-10-24/engine.md b/audit/2026-10-24/engine.md
new file mode 100644
index 00000000..0b3d5d0c
--- /dev/null
+++ b/audit/2026-10-24/engine.md
@@ -0,0 +1,140 @@
+## Engine: media, render, export, core, effects, host playback/preview/posters. main 2606aff1, 2026-10-24
+
+Labels: **Verified** means I read it at the line. **Reasoned** means I followed the code path; nothing was run. **Suspected** means the reason is given. Nothing was built or run: this is a read-only audit, and another agent held the target dir. wgpu-core and wgpu-hal 30.0.1 were read from `~/.cargo/registry`.
+
+Since c41865e these crates have 29 non-merge commits. That is about +5 700 / −1 000 lines.
+- `concat-render/src/gpu.rs` and `concat-effects` changed only in their copyright header.
+- The work is in `concat-media` (decode, pool, audio, samples, encode, feed, ffi, probe, prefetch, svg), `concat-export/src/lib.rs` (+1 054) and `concat-host` (playback +434, media +275).
+
+### 1. The 20 October findings
+
+| Finding | Status | Evidence now |
+|---|---|---|
+| H7 export recovery opens the second device that crashes hybrid laptops | **Unchanged** | `recovered(1)` is still `WgpuCompositor::new()` (`concat-export/src/lib.rs:1612-1624`). It is called from the lost-device loop (`lib.rs:1946-1965`). The sibling still shares the window's error counter (`concat-render/src/gpu.rs:960-963`). |
+| M1 MediaCodec seats first-come, never reclaimed | **Partly, and one part worse** | Better: the export closes idle preview readers before it starts (`concat/src/panes/export.rs:480-490`, 0c86f62), and bin cards decode in software (`concat-host/src/media.rs:925-951`). Unchanged: 3 seats, no priority, no reclaim (`concat-media/src/hardware.rs:268, 320`). Filmstrip `read_through` and `still_at` still ask for hardware (`media.rs:503, 888`). Worse: #505 allows 3 readers per file (`pool.rs:64`), so one file can now hold every seat. |
+| M2 / #448 adjustment layer turns transitions into dissolves | **Fixed** (d8b024f4) | `composite_treated` hands live treatments to `combine_transition`, which grades below and above the incoming lane (`lib.rs:2139-2149, 2219-2238`). The monitor's GPU fast path steps aside when a treatment is live (`lib.rs:2347-2353`). There is a GPU test (`lib.rs:2676+`). Still left: only the first live span gets its shader (`lib.rs:2145`). |
+| M3 GLES export dies after two 2.15 s waits | **Unchanged** | `DEVICE_WAIT` is 30 s (`gpu.rs:296`), passed to `poll` (`gpu.rs:1433-1440`). wgpu-hal is still 30.0.1, so the clamp still applies. |
+| M4 default export: single-pass ABR, `veryfast` | **Unchanged** | `concat/src/panes/export.rs:454-457`. The x264 dictionary is `b`/`maxrate 1.5×`/`bufsize 2×` with no CRF (`encode.rs:831-835`). |
+| N2 chained transitions climb into the next track's lane | **Fixed** (#527) | Lanes are now spread by the longest chain + 1, with cuts resolved in time order (`lib.rs:682-695, 873-879`). Tests cover four transition pairs and a logo over a chain (`lib.rs:2957+, 4193+`). |
+| 12Oct#4 / H1 WARP+FXC compiles unbounded, Cancel deaf | **Partly** | The trigger is narrower: safe start now lasts one build (f26ebf64, window crate). The engine is unchanged: FXC pin at `gpu.rs:861-862`, a synchronous `shader()` at `gpu.rs:2615+`, Cancel checked only per frame (`lib.rs:1801`), no pipeline cache. |
+| Device-lost panic (#475) | **Unchanged in the engine** | No engine commit addresses it. The compositor's handlers log instead of panicking (`gpu.rs:916-934`). 3f6db386 stopped sharing a WARP device with Skia (window crate). That is the one path the repo shows wgpu panicking in its queue. A hardware device shared with Skia is unchanged. #475 is open with no log. |
+| M5 export reader pool fixed at 512 MB on any device | **Unchanged, slightly worse** | `ReaderPool::with_defaults()` (`lib.rs:1754`, `pool.rs:668-671`). Each warm reader now also pins `last` and `ahead` frames outside the byte budget (`pool.rs:374-377`). |
+
+**Tally (9):** 2 fixed, 2 partly, 5 unchanged.
+
+### 2. New findings
+
+#### HIGH
+
+**E-H1. On Android, a safe start leaves the phone with no compositor. The monitor draws nothing and every export fails until the next build.** *Verified path. The phone behaviour is reasoned.*
+- 7df9f8f2 turns safe start on for Android (`concat/src/lib.rs:2643-2645`). Its comment says that on Android `CONCAT_GPU=software` "means OpenGL ES in place of Vulkan (the Skia patch reads it), not the CPU" (`lib.rs:2657-2662`).
+- `concat-render` reads the same variable as "software adapters only" (`gpu.rs:828-833`).
+- wgpu-core keeps only `DeviceType::Cpu` adapters under `force_fallback_adapter` (`wgpu-core-30.0.1/src/instance.rs:511-518`). A GLES adapter counts as `Cpu` only for SwiftShader, llvmpipe or Mesa offscreen (`wgpu-hal-30.0.1/src/gles/adapter.rs:144-155`). A real phone has none of those.
+- Android lends no window device to the engine (`concat/src/platform.rs:671-684` returns `None`). So the export goes to `best_compositor()` and fails with "this device's GPU could not be used to draw with" (`concat-export/src/lib.rs:1573-1592`).
+- The monitor's frames go through `headless()`, which logs "previews have nothing to draw on" (`lib.rs:1643-1660`). Effect cards are skipped (`concat-host/src/cards.rs:398-415`).
+- **Scenario:** the OS kills the app in its first 15 s (low memory, or a swipe from recents), or #507's Vulkan crash happens once. On every launch of that build the window opens on GLES as intended, but the preview is blank and no export can run. This is on the platform with the most downloads.
+- **Fix:** give the window and the compositor separate keys (for example, `CONCAT_WINDOW_GLES`). Or, on Android, have `own_api_first` read the variable as "GL backend first", not "software only".
+
+**E-H2. H7 carried: export recovery opens a fresh HighPerformance device.** Unchanged; see the table.
+
+#### MEDIUM
+
+**E-M1. #462 is fixed only for footage at the timeline's rate. Fast sources still open two or three readers on plain playback.** *Reasoned.*
+- `claim_cost` forgives a reader up to `prefetch::AHEAD` = 8 frames past the target (`pool.rs:335-343`, `prefetch.rs:38`). That 8 is *output* frames of lookahead, but the pool counts in *source* frames at the file's cadence.
+- The race spans the workers' neighbouring *output* frames: 2-4 workers (`prefetch.rs:261`). The table shows where it falls through.
+
+| Source | Source frames per output frame | Output frames the slack covers |
+|---|---|---|
+| 30 fps on a 30 fps timeline | 1 | 8 |
+| 60 fps | 2 | 4 |
+| 120 fps | 4 | 2 |
+| 240 fps slow-motion | 8 | 1 |
+| 4× speed ramp | 4 | 2 |
+
+- bac31a3 makes this more common. `pick_cadence` takes `r_frame_rate` whenever it is above the average (`probe.rs:285-289`), so 1080i MTS/M2TS footage ("29.97 fps, 59.94 tbr") is indexed at the field rate.
+- **Scenario:** a 120 fps phone clip, or a 2-4× ramp, on a 4-worker desktop. Playback opens extra readers, and they decode the same stretch twice.
+- **Fix:** scale the slack by source frames per output frame, or by `workers × step`.
+
+**E-M2. Audio seeks are not checked for landing. With 93c18b72, a late landing becomes silence at the head of the clip.** *Reasoned. The late landing itself is documented by d02b7565's own comment (`decode.rs:1070-1082`).*
+- Video now retreats when a TS/PS seek lands a GOP late or at EOF (`decode.rs:1083+`). The sound side seeks once and trusts it, in both readers: `AudioDecoder` (`samples.rs:209-217`) and the export mix (`audio.rs:759`).
+- Before, a late landing slid the sound early. Now `late` and a negative `lead` pad the overshoot with silence (`samples.rs:94, 247-252`; `audio.rs:383, 785`).
+- A seek that lands at EOF gives a silent clip.
+- No test seeks into a file with both streams: the TS audio fixture is audio-only, and the TS video fixture is video-only.
+- **Scenario:** a camera MTS clip trimmed from the front. In the export and in the preview, its sound starts up to a GOP late, or never.
+- **Fix:** the same landing check and retreat on the audio side. Or seek on the audio stream's own index rather than −1, which TS resolves to video.
+
+**E-M3. The phone's software rerun (0c86f62) may not be able to take the exports most likely to stall.** *Suspected: AOSP limits from memory; needs a device check.*
+- The rerun goes straight to `c2.android.avc/hevc/av1.encoder` (`encode.rs:621-634, 653-662`).
+- As far as I recall, AOSP's C2SoftAvcEnc caps the frame at about 2560 a side and C2SoftHevcEnc at 1080p, both 8-bit. `c2.android.av1.encoder` exists only on newer Android.
+- A 4K or 10-bit HEVC export whose chip stalls renders every frame twice, then fails with "the slower software encoder … failed too".
+- The 'pass loop itself has no end-to-end test. Only `again_in_software` is tested (`lib.rs:3665`).
+
+**E-M4. M1 carried, now with more seat demand per file.** See the table. On Android, a transition within one file, plus a scrub, can take all 3 seats, so the next clip under the playhead decodes on the CPU.
+
+**E-M5. Carried unchanged.** M3 (GLES waits), M4 (default bitrate) and M5 (export pool).
+
+#### LOW
+
+- **L1. A bus placed for a window that starts inside it runs past its end.** `bus_stage` clamps `start` to the window and then adds the full duration (`audio.rs:321-322`). Playback windows start at the earliest layer of a group, so this cannot happen today. Latent for any future `mix_pcm` caller. Also, each bus chain runs over the whole timeline, gated to silence outside its span: an `asplit=3` per bus, so the cost grows with timeline length, not span length.
+- **L2. Posters can fill the worker kept for playback.** `Poster` is not background work (`prefetch.rs:83-87`), so a big import's posters can take every worker. A poster on a slow volume (a network share, an iCloud placeholder, a 4K HEVC probe on a phone) then makes playback wait for one to finish. The measured gains (7b2d30e) justify the priority; a cap of `workers - 1` would keep both.
+- **L3. Damaged-packet skipping never trips on a stream that alternates.** A good send resets `damaged` (`feed.rs:108-117`). A stream whose every packet sends fine and decodes to `InvalidData` is read to EOF under the reader's lock, where it used to fail at once. The cost is time, not a hang.
+- **L4. The racing claim re-decodes frames.** A reader claimed from behind keeps its furthest `aim` (`pool.rs:561-565`) after `pull` has sought it back. The next claim then rolls up to 8 frames it already decoded. The cost is CPU only.
+- **L5. SVG drawings are cached in a shared, predictable temp path** (`svg.rs` `drawing_of`: `temp_dir()/concat-svg/<hash>.png`). On multi-user Linux, another user can plant the PNG. `load_system_fonts` runs on every first draw.
+- **L6. Projects imported before b8337561 may keep their squeezed sizes.** They store the 1440-wide media size, and export `decode_sizes` come from the clip's stored dimensions. Unverified whether the app probes again on open.
+
+### 3. Commits reviewed
+
+| Commit | Verdict |
+|---|---|
+| #531 fe2be741 `timeline_extent` | Correct. It uses the same `quantise_at` + `quantise` as `build_timeline` (`resolve.rs:291-292`, `lib.rs:1281-1287`), and the mix takes the same end. Tested. |
+| #531 167a7331 borrowed plan | Correct. It clones only when a treatment is live (`lib.rs:2150-2154`), and the retry after a lost device redraws from the same borrow. |
+| #527 `lane_spacing` | Correct. A chain of n cuts climbs n lanes with n+1 spacing (`lib.rs:873-879`). The sound-layer ordering by track survives the multiply. Tests cover four transition pairs. |
+| #505 + 2b6eb23 readers per key | Sound locking: claim under the short lock, decode under the reader's own; no nested locks, so no deadlock. The LRU is a `Vec`, which is fine at 8. The units gap is E-M1; the seat cost is E-M4. |
+| 0c86f62 software rerun | The loop is bounded (`software_only` gates a second rerun, `lib.rs:2015`). Decoders are dropped each pass, the encoder is dropped before the reopen, and progress resets in the window. Open: E-M3. |
+| d02b7565 TS/PS retreat ladder | Bounded: steps of 1, 4, 16 … s, then `seek_top` clears `landing` (`decode.rs:1083-1115`). Codec flush on retreat is correct after EOF. Byte seek only for `AVFMT_TS_DISCONT`. Good. |
+| b8337561 SAR | Matches `av_guess_sample_aspect_ratio` precedence, with a sane band (`ffi.rs` `sample_aspect`). The stretch comes before rotation. Tested with an HDV fixture. See L6. |
+| 39ebb75e ms-timestamp tolerance | Capped at a quarter of an output frame (`decode.rs:1427`), so the next frame is never pulled in early. Tested with MKV. |
+| 1dac19b7 held frame before target | Holds at most one undecoded hardware frame (`short` or `stashed`, `decode.rs:1236-1249`), well inside D3D11VA's spare surfaces, so no surface exhaustion. Frames handed on are downloaded first. |
+| bac31a3 VFR cadence | Correct for screen recordings. Over-eager for interlaced and jittery files (E-M1), but `held_at` keeps the picture right. |
+| 7c113230 damaged packets | Right direction, bounded at 64 in a row. See L3. |
+| 93c18b72 one origin per file | Consistent across picture, preview sound, waveform, transcript and export mix (`ffi::origin` in `decode.rs`, `samples.rs:200`, `audio.rs:751`). The audio seek gap is E-M2. |
+| Poster/Waveform priorities | Sound ordering, with tests. See L2. |
+| ccc6131c `card_still` | Good: software, one keyframe, with exact-frame fallbacks for black, unreadable or later frames. Tested against a scene-cut stream. |
+| 7dc91a92 filmstrip keyframes | Good: it keeps the result only if every tile came from its own slice, and otherwise falls back to the exact pass. It decodes on the platform's video hardware (`keyframes_only` keeps it off MediaCodec). |
+| 2606aff1 sound layers | Graph is correct. The dry path minus the ramped dry plus the ramped wet is exact outside the span (a test checks it sample for sample). The playback span is the export's own `run_mix`. Clips go silent only once the span is cached (`playback.rs:805-809`). Cache is budgeted. See L1. |
+| e353fe57 WAV channel layout | Correct. It is applied before the buffer source is described, in both readers. |
+
+### 4. Numbers
+
+| Crate | `#[test]` now (20 Oct) | Ignored |
+|---|---|---|
+| media | 143 (113) | 1 |
+| render | 101 (101) | 0 |
+| export | 46 (38) | 0 |
+| core | 64 (60) | 0 |
+| effects | 72 (72) | 0 |
+
+Big files: export `lib.rs` 4 358 (+588 since 20 Oct), `gpu.rs` 3 695, `encode.rs` 2 813, `decode.rs` 2 488, `pool.rs` 1 872, `audio.rs` 1 701, `playback.rs` 1 779.
+
+### 5. Scores
+
+| Aspect | 20 Oct | Now | Why |
+|---|---|---|---|
+| Code quality | 7 | 7 | The #457 fixes are careful and bounded (retreat ladder, damage counter, slack capped at a quarter frame). But `CONCAT_GPU` means two different things in two crates (E-H1), and `claim_cost` mixes units (E-M1). |
+| Design architecture | 7 | 6.5 ↓ | Readers routed by aim and a sound-layer bus over the sum are good designs. Seats still have no priority, recovery still opens a second device, and safe start has no single owner. |
+| Performance | 6 | 7 ↑ | Measured, large wins: card pictures 8-15 s → 0.1-0.2 s, phone filmstrips 10-13 s → 0.15 s, and no plan clone per frame. Offset by extra readers for fast sources and a full-timeline cost per bus. |
+| Scalability | 5 | 5 | Fixed 512 MB export pool; 3 readers per file × 8; frames pinned outside the budget; the audio mix still holds the timeline (N4). |
+| Testing | 6.5 | 7 ↑ | A real fixture for each #457 format (TS, MKV, HDV, VFR, late picture and late sound), a chained-transition test and a sound-bus sample test. Missing: Android safe start with the compositor, audio seek in a muxed TS, fast-source claims, the 'pass loop end-to-end. The device test is still `#[ignore]`. |
+
+### 6. Must fix before a release
+
+1. **E-H1.** One launch that dies early leaves an Android build with no preview and no export. It is a one-line separation of meaning, on the priority platform.
+2. **E-M2 needs a check.** Export a camera MTS clip trimmed from the front, and listen to the first second.
+3. **E-M3 needs a device check.** A 4K export through `c2.android.avc.encoder`.
+
+### 7. Not checked
+
+- MediaCodec and C2 software encoder behaviour on real phones; I did not test their size limits.
+- Whether the phone shows the Settings GPU switch that would clear a safe start.
+- #475's real cause: there is no log in the issue.
+- Whether old projects probe HDV clips again (L6).
diff --git a/audit/2026-10-24/host-api-security.md b/audit/2026-10-24/host-api-security.md
new file mode 100644
index 00000000..b49c60df
--- /dev/null
+++ b/audit/2026-10-24/host-api-security.md
@@ -0,0 +1,152 @@
+## host, server, API, MCP, CLI, project, speech, vision, and security across the codebase: read at 2606aff1 (2026-10-24)
+
+Read-only. Nothing built or run (another agent holds the shared target dir). "Verified" means I read it at the line given in the current tree. "Reasoned" means inferred from code I read. "Unverified" means I could not check it. Outside checks: `gh` (read-only), the Hugging Face and Gradle checksum endpoints, the upstream whisper.cpp commit through the GitHub API.
+
+Baseline is c41865e (20 Oct). Since then, 102 commits. In this area most of the diff is the copyright-header rename. Real changes: concat-server `json.rs` +76 and `lib.rs` +53 (the #431 stop), concat-api only `message.rs` docs, concat-mcp `protocol.rs`/`tools.rs` (+178), concat-cli and concat-vision headers only, concat-speech `transcribe.rs` +132 and an 885 KB model, concat-host `net.rs` (new, 82), `projects.rs` +40, concat-project +1 000 (speed curve, MatchClip, sound layers), concat-mobile (new, 1 321 lines), `android-native/` (new Gradle spike).
+
+Short version: two things got better. Prefs are now atomic and 0600, and captions are timed from Silero VAD plus the upstream whisper fix. One thing got worse: the updater now trusts any root certificate in the system store, so the one barrier it had, Mozilla's root list, is gone. Every server and MCP finding from 20 Oct is untouched.
+
+### Numbers (`#[test]` count, `unsafe` count, `.unwrap()` count, all `.rs`)
+| crate | lines | tests | unsafe | unwrap |
+|---|---|---|---|---|
+| concat-host | 18 547 (+~1 000) | 137 (+12) | 46 | 61 |
+| concat-server | 2 598 (+85) | 28 (+1) | 0 | 0 |
+| concat-api | 2 973 | 26 | 0 | 0 |
+| concat-mcp | 3 199 (+158) | 13 (+2) | 0 | 0 |
+| concat-cli | 520 | 3 | 0 | 0 |
+| concat-project | 13 777 (+878) | 165 (+15) | 0 | 41 |
+| concat-speech | 3 635 (+131) | 24 (+2, 1 ignored) | 5 | 10 |
+| concat-vision | 5 498 | 59 | 0 | 5 |
+| concat-mobile | 1 321 (new) | 2 | 10 | 0 |
+
+---
+
+### 1. Status of the 20 Oct findings in this area
+
+| 20 Oct item | Status | Evidence (current tree) |
+|---|---|---|
+| H5 / S1: updater checks a digest from the same release | **WORSE** | Still no signature: `updates.rs:375-414` reads `sha256` from `manifest.json`, which `manifest_url` (`:416-418`) fetches from the same release, and `download_package` (`:440-473`) checks against it. `release.yml:151` writes the manifest unsigned, with no attestation. Quarantine is still stripped (`updates.rs:584-586`). New: 46af73c sends `get_text` (`:671`) and the package download through `net::agent` (`net.rs:23-29`), which trusts the system root store as well (`net.rs:56-60`). See H1 |
+| M1: shared auth-throttle bucket | UNCHANGED | Every TCP peer is keyed by IP (`json.rs:95`), so all local callers share one key. Any bad first line counts as a failure (`:269`, `handshake :405-411`, which includes `GET / HTTP/1.1`). Refusal comes before the token is read (`:229`). `throttle.rs:6-7` still says a minted token needs no throttle, and the code throttles it anyway |
+| M2: agent-opened projects stay locked | UNCHANGED | A connection end only unsubscribes (`json.rs:343`). `Window::finish` is still the no-op default (`concat-mcp/src/backend.rs:47`). Claims are released only by `close` or by `Drop for Api` (`api lib.rs:578-591, 1067-1074`) |
+| M3: MCP hints understate writes | UNCHANGED | `preview_frame` is `read_only: true, destructive: false` (`tools.rs:462-468`), yet with `output` it `File::create`s any path under the roots (`api lib.rs:811, 1148-1160`). `export_start` is not marked destructive and overwrites (`:665`). `docs/MCP.md:47` still sells "only under your home folder" as the safety property |
+| M4: lock opens unguarded on I/O error | UNCHANGED | `api lib.rs:186-188` (`LockError::Io(_) => None`). Windows `same_file` is still `path.exists()` (`lock.rs:261-262`) |
+| M5: shipped `concat --mcp` never run in CI; dead `catch_unwind` | UNCHANGED | No `--mcp` in `.github/` or `scripts/ci`. `panic = "abort"` (`src/Cargo.toml:127`); the comment at `:124-126` now admits it |
+| Prefs world-readable / non-atomic | **FIXED** | 4b1b2b05. Staged `settings.json.<pid>.part` → rename (`concat/src/prefs.rs:352-363`). Created 0600 on Unix (`:391-401`), test `:453-457`. A bad value now costs only itself (`salvage :368-387`), and the original is kept as `.unreadable` (`:337`). Small hole: a stale `.part` left by a crash keeps its old mode, because `mode()` applies only on create (Low) |
+| Downloads without size caps | UNCHANGED | `models.rs:224-305` writes until EOF with no cap against `estimate`. Sounds use the same path |
+| Phone FFmpeg / sherpa pulled unverified | UNCHANGED | `src/scripts/ffmpeg-mobile.sh:39`, `sherpa-mobile.sh:42, 60`: `curl` with no digest. Desktop is digest-pinned (`scripts/ci/fetch.sh:66-79`) |
+| Model mirror 404 | UNCHANGED | `gh api .../releases/tags/models-v1` returns 404 today, and `models.yml` has no runs. `THIRD_PARTY_NOTICES.md:24-26` still says Concat "mirrors them and downloads them on demand from that mirror" |
+| H8: captions start early (#438) | **FIXED in code**, unverified on audio | Two layers. The upstream whisper.cpp parser fix (`patches/whisper-rs-sys/.../whisper.cpp:7659-7687`, PR #528). Silero VAD spans clamp every segment (`transcribe.rs:237-300, 602, 625`). Only `on_speech` is unit-tested (synthetic spans). No test runs the bundled VAD on audio, though it ships in-tree and needs no download. #438 is still open, correctly, with no reporter confirmation |
+| L1 Window backend retries non-idempotent calls | UNCHANGED | `backend.rs` header-only diff |
+| L2 token estimator / IPv6 eviction | UNCHANGED | `token.rs:18, 53-72` |
+| L4 Recents non-atomic | UNCHANGED | `projects.rs:545-553` (`fs::write`) |
+| L10 sounds trust a same-manifest digest | UNCHANGED, now wider (H1) | `sounds.rs:382-395` |
+| L11 doc drift | PARTLY | `THIRD_PARTY_NOTICES.md:6` still lists libavdevice, which af907228 dropped. `dirs.rs`, `protocol.rs`, `prefs.rs` and the other L11 comments are unchanged apart from headers |
+| S17 Unix socket perms from umask | UNCHANGED | `server lib.rs:272-275` |
+| 11Oct#10 API create unchecked | UNCHANGED | no diff in `api lib.rs` beyond the header |
+| N6 (project): transition params bypass NaN guard | **FIXED** | `commands/mod.rs:1136` now checks `tr.params` |
+
+Tally, 20 items: 3 fixed, 1 partly, 1 worse, 15 unchanged. The server, API and MCP list got no work since 20 Oct.
+
+---
+
+### 2. New findings
+
+**HIGH**
+
+**H1. The updater now installs whatever a TLS interceptor serves.** Verified (code); unreleased (46af73c, after v0.2.8).
+- `net::agent` trusts webpki-roots **plus** every root in the OS store (`net.rs:52-62`). rustls-native-certs also honours `SSL_CERT_FILE` and `SSL_CERT_DIR`.
+- The updater has no signature. The digest it checks comes from a manifest fetched over the same connection (`updates.rs:416-473`). So before this commit, the only thing standing between a network attacker and code execution was "the certificate must chain to Mozilla's list". That is now "to any root this machine trusts".
+- The commit's own motivating case is exactly the attacker: "a proxy or a GitHub accelerator in China, an antivirus that scans HTTPS, a company's gateway" (`net.rs:9-13`). Each of these re-signs github.com and can now swap `manifest.json` and the package. The updater then runs `setup.exe` / `msiexec` with no Authenticode check (`updates.rs:489-505`), or swaps the `.app` and strips quarantine (`:584`).
+- Models are fine: their digests are compiled in (`models.rs:311-330`). Sounds are not: digest and file come from one fetched manifest (L10), so an interceptor can feed FFmpeg's parsers arbitrary audio.
+- Fix before the next release, in order of cost:
+ 1. Give `updates.rs` (and `sounds.rs`) a webpki-only agent. This restores the 0.2.8 posture; it is one function.
+ 2. Sign the manifest with an ed25519/minisign key whose public half is compiled into the app, and check the signature before reading the digest.
+ 3. On Windows, check WinVerifyTrust on the installer once signing exists.
+- Keep system roots for model downloads only. That is where #517's users actually failed, and the digests make the wider trust harmless there.
+
+**MEDIUM**
+
+**M6. A cross-project paste still carries the other project's media ids, so Enhance-off and image backgrounds play the wrong file.** Verified.
+- 1d04d147 remaps `clip.media_id` by path (`studio.rs:8288-8320`). `MatchClip` then copies the rest of the source clip wholesale (`commands/properties.rs:330-360`). That includes `enhanced { original, copy }` (`model.rs:1310-1317`) and `background: Image { media_id }` (`model.rs:731-735`), and both are bin ids of the *source* project.
+- Failure: copy an enhanced clip in project A and paste it in B. Toggle Enhance off. `SetClipEnhanced` sets `media_id = link.original` (`commands/clips.rs:498-515`), which is A's `m3`. In B that is another file or none. An image background draws B's `m4` or nothing.
+- The commit message says "a paste into another project plays the right file", which is true only until either of these is touched.
+- Fix: in `held_here`, remap `enhanced.original`/`copy` and `Background::Image` by path the same way, or clear them when the target has no such file. Add a test with both fields.
+
+**M7. The VAD step runs blind: no cancel, no progress, a fixed-name `.part`.** Verified; timings reasoned.
+- `speech_spans` runs over the whole window before `state.full` (`transcribe.rs:602`). Whisper's abort callback is not installed for it, and progress sits at the first tick. On an hour of audio Silero takes on the order of tens of seconds on CPU (unmeasured), and Cancel does nothing until whisper starts.
+- `vad_model` writes `ggml-silero-v6.2.0.part` (`path.with_extension("part")`, `:288-300`), the same name in every process. Two windows captioning at once truncate each other's file, and one of them may load a half-written model. That run falls back to whisper's times with only a log line.
+- "Whole" is decided by length alone (`:290-291`).
+- Fix: check `cancel` around the VAD call (or run it on the abort hook). Use a pid/nonce `.part` as `prefs.rs:360` now does. Compare the bytes once, or a digest, rather than the length.
+
+**M8. Forward compatibility: a project with a sound layer loses it on the first autosave in 0.2.8.** Verified against `v0.2.8:model.rs:1652-1655`.
+- 0.2.8 reads the unknown kind `"soundLayer"` as `Video` (`wire::clip_kind`). `settle_clip` then drops it, because a Video with an empty `media_id` has no media (`doc.rs:264-272`). The next save writes the project without it. `DOCUMENT_VERSION` stays 1 (`doc.rs:41`).
+- The model doc (`model.rs:67-77`) calls this "leaves it out as a clip it cannot play". That holds until the autosave.
+- Phones lag desktop releases, and a project synced between a phone on 0.2.8 and a desktop on main is the realistic case. Same class as 20 Oct N7, now losing whole clips.
+- Fix: bump a minor document marker that older builds refuse to *save* over (open read-only), or carry unknown clips through as opaque `extra`.
+
+**LOW**
+- **L13.** On Windows, cross-project paste looks a file up by exact path (`studio.rs:8297`), while `AddMedia` dedups case- and slash-insensitively (`commands/media.rs:11-27, 37-41`). Take `C:\v\a.mp4` against `c:/v/A.mp4`: `held_here` finds nothing, `AddMedia` returns `Outcome::default()`, `apply_in` gives `None`, and that clip is silently not pasted. Use `same_path` in both places.
+- **L14.** concat-mobile (a spike, not shipped). `Engine::attach_surface(window: u64)` is a safe UniFFI export that turns a Kotlin integer into `NativeWindow::from_ptr` (`render.rs:469-474`): an unsound boundary, with only null checked. `Engine::new` calls `set_var` (`lib.rs:141-145`) under a SAFETY note that fails on a second construction, because the first engine's threads are running. `ANativeWindow_fromSurface`'s reference leaks if `attach_surface` is never called (`jni.rs:47-55`). `gradle-wrapper.properties` has no `distributionSha256Sum`; the committed wrapper jar does match Gradle's published 8.12 checksum (verified).
+- **L15.** `platform.rs:451-454` calls `set_var("CONCAT_GPU")` after the GPU probe, under a SAFETY claim that the process is single-threaded. On Linux, llvmpipe and other Mesa drivers have started threads by then (reasoned). This is a Unix setenv race; Windows is unaffected.
+- **L16.** SVG import (f5664c83, `concat-media/src/svg.rs`) uses `usvg::Options::default()`, so an SVG's `<image href="/abs/path">` is read from disk and drawn into the still (reasoned from usvg's default resolver). A downloaded SVG can embed a local picture into an export the user later shares. Set `resources_dir: None` with a resolver that refuses file hrefs, or allow `data:` only.
+- **L17.** `concat-speech/Cargo.toml:18` still depends on `ureq`, which nothing in the crate uses. `THIRD_PARTY_NOTICES.md` names the MIT licences of whisper.cpp and Silero but reproduces neither text, which MIT's notice condition asks for in binary distributions.
+- **L18.** The MCP stage names are string literals shared across three crates (`concat-export lib.rs:1250-1257, 1765, 2000`, `concat-mcp tools.rs:43-52`, `api message.rs:429-433`). A rename in the exporter silently drops progress to stage 0. Export a constant set from concat-export.
+- **L19.** `ci.yml`/`mobile.yml`: dependabot moved `dtolnay/rust-toolchain` to 683b588 but the comment still reads "2026-10-09". Workflows other than `models.yml`, `nix.yml` and `release.yml` still have no `permissions:` block (20 Oct L2).
+- **L20.** The MCP line reader splits a line over 16 MiB into chunks, each parsed on its own (`protocol.rs:44, 155`). This input comes from the client, so it is not a security issue, but the tail of a long line can run as a request. Discard to the newline instead.
+
+---
+
+### 3. Commits since c41865e, area view (independent review)
+
+- **2922614c VAD (#438).** Correct in shape.
+ - The bundled model is the real one: sha256 `2aa269b7…6987` matches Hugging Face's `x-linked-etag` for `ggml-org/whisper-vad/ggml-silero-v6.2.0.bin` (verified), and the size, 885 098 bytes, matches.
+ - Centisecond units are right: whisper-rs 0.16 `whisper_vad.rs:261-301` documents `start`/`end` in centiseconds.
+ - `on_speech` cannot invert a span for sorted, valid spans, and falls back to whisper's times when nothing overlaps.
+ - `include_bytes!` links only into `concat` (the one crate that depends on concat-speech).
+ - Holes: M7, plus the missing audio-level test.
+ - Notices entry present (`THIRD_PARTY_NOTICES.md:29-33`).
+- **ea225312 / f1a5f4d3 whisper.cpp patch (#528).** Byte-identical to upstream c33c5618 (whisper.cpp #2279, 2026-05-10, after 1.8.3), compared through the GitHub API. Documented in `patches/README.md:67-88`, including a note to drop it on upgrade. Good contributor work.
+- **46af73c TLS.** Well reasoned for downloads, and the shipped roots are always kept (test `net.rs:72-81`). It applies one trust policy to code and to data alike (H1). No duplicate TLS stack (one rustls 0.23.45, ring 0.17.14; `Cargo.lock`).
+- **1d04d147 MatchClip.**
+ - The right idea: copy the whole clip, keep only the place. `#[serde(skip)]` keeps it off the wire, and the test proves `{"op":"matchClip"}` is refused (`lib.rs:617-618`).
+ - Nothing serializes a `Command` outside a test (grep), so a skipped variant cannot panic a serializer.
+ - Holes: M6 and L13.
+- **2606aff1 sound layer.**
+ - The serde default `sound: false` keeps old API callers working, and old projects read unchanged.
+ - The export chain is kind-filtered: a picture package on a sound layer yields an empty chain, and `audio_buses` (`concat-export lib.rs:1353-1370`) drops it. Strength is clamped to 0-1.
+ - The MCP description of `effect_layer_add` is updated (`tools.rs:391`).
+ - Hole: M8.
+- **77ce50bb concat-mobile spike.** Narrow API, every unsafe has a SAFETY note, and it is not built by any workflow (grep), so it ships to nobody. L14.
+- **4ffa3212 / server stop (#431).** The stop now cancels jobs, joins the API thread, then drains each outbox before hanging up (`server lib.rs:364-370`, `json.rs:298-311`), with a test (`json.rs:796-843`). The drain is sequential across connections, up to 10 s each, but the window bounds it with `EXPORT_PATIENCE` on its own thread (`concat/src/lib.rs:2598-2611`). Fine.
+- **3dd2ee95 / 1a2c71ff / 9f465da4 MCP.** Non-UTF-8 lines answered and the session kept (test `protocol.rs:523-534`). Stage mapping (L18). `ttlMs`/`cacheScope` on `server/discover` and `tools/list` per the 2026-07-28 draft: still **unverified** against the spec (past my knowledge, as on 20 Oct).
+- **e737cf89 device names.** Correct for CON/PRN/AUX/NUL/COM0-9/LPT0-9 and the superscript ports; it misses `CONIN$`/`CONOUT$` (trivial).
+- **4b1b2b05 prefs.** See §1.
+- **f312433c CI.** fmt, locales and models are now in their own job, with `if: !cancelled()` (`ci.yml:35-66`). This answers the 20 Oct "gates hide clippy" finding.
+
+### 4. Security pass
+
+| Surface | Finding |
+|---|---|
+| Port 7420 from a web page | A page cannot drive the API: the first line must be a JSON `auth` (`json.rs:398-422`), and a browser's first line is `GET`/`POST … HTTP/1.1`. CORS and Origin do not apply (not HTTP), and DNS rebinding gains nothing without the token. What a page *can* do is M1: four `fetch()`es lock every local caller out for up to 60 s per strike, indefinitely. Unchanged |
+| Tokens | Not logged anywhere (grep of `log::*!` in server, MCP, settings). Discovery file 0600 (20 Oct); prefs now 0600 (fixed). Windows relies on profile ACLs (unverified) |
+| Path traversal | Writes are confined by `writable()`: `..` refused, symlinks resolved on the existing prefix (`api lib.rs:379-412, 1108-1130`). Opens are unconfined by design, and `project_open` + `project_save` still write `concat.json` and `concat.lock` into any existing project folder (M3). Package and sound names are sanitised (`updates.rs:396-401`, `sounds.rs`), and `folder_name` now also escapes device names |
+| Updater | H1 (worse than 20 Oct): no signature, no attestation, same-release digest, system roots trusted, quarantine stripped |
+| Unsafe added since 20 Oct | `concat-media decode.rs` seek_top and `ffi.rs` sample_aspect: sound (plain field reads on open handles). concat-mobile: 10 (L14). `platform.rs:454` `set_var` (L15). No new unsafe in my crates outside the spike |
+| Secrets in repo | None. Grep for AWS, GitHub, Slack, HF, OpenAI and Google key shapes and private-key headers across tracked files came back empty, and there are no keystores or `.env` files |
+| Supply chain | Desktop prebuilts digest-pinned (`deps.lock`, `fetch.sh`). Phone FFmpeg and sherpa unpinned (unchanged). Actions pinned by SHA. Gradle wrapper jar genuine but distribution unpinned (spike). Homebrew on macOS builds unpinned (20 Oct, not re-checked). Silero model verified against upstream |
+
+### 5. Scores (this area)
+| Aspect | 10-20 | Now | One line of evidence |
+|---|---|---|---|
+| Code quality | 8 | 7 ↓ | Careful new code with tests (stop drain, NUL names, prefs salvage), but MatchClip copies project-local ids (M6) and the VAD `.part` repeats the shared-scratch mistake fixed in saves on 12 Oct (M7) |
+| Design architecture | 7 | 6 ↓ | `net.rs` as the single network choke point is right, but it gives code and data one trust policy (H1). Claims are still not tied to sessions (M2), the lock still fails open (M4), and a new clip kind drops on downgrade (M8) |
+| Security | 5 | 4 ↓ | Prefs fixed (+). Updater trust widened (H1). M1, M3, M4, the size caps and the phone deps all unchanged |
+| Testing | 7 | 6 ↓ | Every commit carries a test, but the VAD is never run on audio although it ships in-tree, there is no `--mcp` smoke, no cross-project paste test with enhance or background, and the TLS test only checks that a config builds |
+| Documentation (API/MCP) | 6 | 6 → | `patches/README.md` and the stage docs are exemplary. `MCP.md:47` still oversells "home", `THIRD_PARTY_NOTICES.md` claims a mirror that 404s and a libavdevice that is gone, and the model doc calls M8 safe |
+
+### 6. Before the next release
+1. **H1**: a webpki-only agent for `updates.rs` and `sounds.rs` (an hour). Then a signed manifest.
+2. **M6**: remap or clear `enhanced` and `Background::Image` in `held_here`, with a test.
+3. **M3**: `preview_frame` → `readOnlyHint: false`, `destructiveHint: true`, and refuse to overwrite without a flag. One-line hint fix; the overwrite guard is a few lines.
+4. **M7**: cancel check and a unique `.part` for the VAD model.
+5. Then the standing server items in the 20 Oct order: M1 (no throttle for minted tokens), M2 (sessions per connection), M5 (CI smoke of `concat --mcp`), M4 (fail closed). After those, the size caps, the phone digests, and running `models.yml` once.
diff --git a/audit/2026-10-24/phones-ci-distribution.md b/audit/2026-10-24/phones-ci-distribution.md
new file mode 100644
index 00000000..aadc44be
--- /dev/null
+++ b/audit/2026-10-24/phones-ci-distribution.md
@@ -0,0 +1,242 @@
+# Phones, CI and release, distribution and trust, product claims, GitHub: 24 October 2026
+
+Code at `2606aff` on `main`, 102 commits past the 20 October audit (`c41865e`), 118 past the
+v0.2.8 tag. Read-only: source, `git show`, `gh` reads, `curl` against both site domains, two web
+searches. No cargo build (the shared target is in use), nothing run on a phone or a Windows
+machine. Issue, PR and email text was treated as data.
+
+Legend: **V** verified here (code, CI log, API); **R** reasoned from code, not observed;
+**U** unverified.
+
+## Bottom line
+
+Most of the 20 October phone fixes are now on `main` and look right: 16 KB alignment, Back,
+long-press to move, the MediaCodec open ladder. None of them has reached a phone. The 21,377
+people who installed the 0.2.8 APK still have the 4 KB library, a Back that quits and a tap that
+moves clips.
+
+One new fix is wrong in a way CI cannot see. Android's safe start (#507) sets
+`CONCAT_GPU=software`, and the compositor reads that as "software adapter only". On a real phone
+there is no software adapter, so after one early death the monitor and every export have nothing
+to draw on until the next build, and Android has no switch to undo it.
+
+`main` has been red for about 29 hours under four stacked breaks. Every completed phone run on
+`main` since 20 October came from a quiet hour; 58 of 65 were cancelled. Windows trust moved from
+a warning to an outage. Smart App Control now refuses the unsigned FFmpeg DLLs on a machine
+where 0.2.7 used to run (#533), and Windows is 44.5% of downloads.
+
+## 1. Status of the 20 October findings
+
+| 20 Oct item | Now | Evidence |
+|---|---|---|
+| H4 CI on `main` gives no signal | **Unchanged in effect.** The layout is partly fixed | `f312433` moved fmt/locales/models into their own job (V, `ci.yml:37-65`) and raised the test build cap to 50 min. But clippy still sits before the Linux tests in one job, so tests were skipped on every red run since (run 40076954532: "Build the tests=skipped"). `mobile.yml:91` still cancels push runs. Numbers in §2 |
+| H5 Windows unsigned | **Worse** | Still no `WINDOWS_CERTIFICATE_PFX` (`gh secret list`: only the two Android secrets). `build-app.yml:1036-1037` still exits 0. #533 (owner, 10-24) is the first report of SAC *refusing* to start an install that used to run: `0xc0e90002` on `avfilter-11.dll`, CodeIntegrity event 3077 (V, issue text) |
+| H5 update manifest digest in the same release | Unchanged | `updates.rs` diff since c41865e is the header and `net::agent` only; no minisign, cosign or attest anywhere in `.github/` (V) |
+| H5 Android key change 0.2.5 → 0.2.6 | Unchanged, no note posted | #526's commenter suggests uninstalling first, which fits a signer conflict (U) |
+| H5 Android developer registration | Unchanged | No mention in the repo outside `reports/` (V). Enforcement began 30 Sept in BR/ID/SG/TH. Phase 1 reportedly covers installs from participating stores; global in 2027 ([TechEnclave](https://techenclave.com/news/items/android-developer-verification-begins-in-four-markets-not-india-yet), [Google](https://developer.android.com/developer-verification/guides?authuser=3), [TechRepublic](https://techrepublic.com/article/news-google-android-sideloading-24-hour-wait)) (U on scope) |
+| SECURITY.md channel disabled | **Fixed** | `private-vulnerability-reporting` → `{"enabled":true}` (V) |
+| H6 macOS floor vs shipped dylibs | **Partly. The danger is now proven** | `af90722` drops libavdevice. #532 (10-24, macOS 13.7.8, 0.2.8) died on exactly that: `_AVCaptureDeviceTypeContinuityCamera`, "libavdevice built for macOS 14.0" (V, crash text). The rest of the Homebrew set is still built for 14/15. `check-min-os.sh` still only warns (`build-app.yml:647-654`) |
+| 16 KB alignment | **Fixed on `main`, not released** | `mobile-env.sh:70` sets `max-page-size=16384`. `check-16k-pages.py` fails on any `lib/*.so` LOAD align < 16384. Phone run 39902059739: all four libs 16384 (V). The libraries are packed compressed (`android-native/app/build.gradle.kts` note), so zip alignment is moot |
+| H2 Android Back quits / no save on exit | **Partly** | `0aa1a50`: Back = Escape over the desk sheets; with nothing on top the project is saved and the app goes to the launch screen; the loop end saves (V). Missed: the phone editor's own sheet (M1). No save on Pause or Stop, but autosave is 1.5 s (`studio.rs:2903`), so the loss is small |
+| H3 tap moves a clip | **Mostly fixed** | `b467155`: under a finger a press is a tap, a swipe or a 400 ms hold, with 8 px slop (V, `lanes.slint`). The phone shell's own controls are 44 px. The pinch and hold edge cases are in Low |
+| Toast / tooltip insets | Unchanged | `ToastLayer` covers the full window with no inset (`app.slint:995-1000`) |
+| Soft-keyboard inset | Unchanged | No keyboard-height handling anywhere in `ui/` or `src/` (V, grep) |
+| RTL mirroring | Unchanged | No layout mirroring for ar/he/fa/ur (V, grep) |
+| Accessibility names | Unchanged | 7 `accessible-label` in all of `ui/`, the same as at c41865e (V) |
+| Microphone permission dead end | Unchanged (not re-traced) | `ConcatFiles.java` only had its header touched |
+| Phone smoke accepts a system ANR | Unchanged | `smoke-android.sh` untouched |
+| Phone FFmpeg/sherpa downloads without digest | Unchanged | `ffmpeg-mobile.sh`, `sherpa-mobile.sh` untouched; no sha256 in either (V) |
+| Model mirror `models-v1` | Unchanged | Release not found; `models.yml` has never run (V) |
+| Website stale + checksum mismatch (#458) | **Fixed for 0.2.8; structurally open** | Both domains serve 81× "0.2.8" and a `v0.2.8/SHA256SUMS` link (V, curl). Links come from the release manifest *at build time* (`website/src/lib/releases.ts`). Deploys run only on a push to the website repo (`deploy-pages.yml`), so the next release leaves the site stale again. #458 is still open |
+| README platform ticks | Unchanged | `README.md:82-86`: ✅ Windows, Android, iOS. "4K scrubs smoothly" (`:67`). The SAC refusal is not mentioned |
+| 0.2.8 notes: false "Q and E cut" | Unchanged | Release body line 55 (V) |
+| 0.2.8 notes: "Intel Macs open on macOS 12 again" | **Unchanged and now disproven** | Line 66, against #532 on 13.7.8. 0.2.7's notes were corrected after the fact; 0.2.8's were not |
+| Dependabot cargo group red | **Worse** | #411, then #519 (closed by dependabot 10-24, superseded), now #534: 9 of 11 checks fail (V). See M3 |
+
+## 2. CI on `main` since 20 October (V, `gh run list --branch main`, 148 runs from 10-20 10:07Z)
+
+| Workflow | Runs | Success | Failure | Cancelled | Running |
+|---|--:|--:|--:|--:|--:|
+| CI (push) | 69 | 4 | 14 | 49 | 2 |
+| Build for phones (push) | 65 | 6 | 0 | 58 | 1 |
+| Nix | 5 | 5 | | | |
+
+- **The last green CI** is `f312433` (10-23 03:11Z), the CI change itself. Every CI run on `main`
+ that completed after it failed.
+- **Why it is red now**, four breaks stacked. Each landed in a burst of pushes that cancelled the
+ run that would have caught it:
+ 1. Locales: `start.safeMode` and `start.safeModePhone` went unseen by the scanner because
+ `7df9f8f` picks the key through `cfg!`. Red from `4b1b2b0`, fixed in `fba598a` a day later.
+ 2. Locales: `22639548` removed `clipInspector.bypass`, `.enable` and `.favourites` without
+ regenerating `en.json`.
+ 3. Fmt: the spike `77ce50b` (`concat-mobile/src/render.rs` ×10, `lib.rs`), `transcribe.rs`
+ (from the #528 merge) and `pool.rs`.
+ 4. Clippy: `needless_borrow` at `concat-host/src/cards.rs:516` (`5296d98`), plus lints in
+ `concat-speech/src/transcribe.rs:280` and `patches/whisper-rs-sys/build.rs:154`.
+ The Windows tests do run and pass (run 40076954532). The Linux tests are skipped behind clippy.
+- **Pace:** 20 pushes between 06:58Z and 09:07Z on 10-24. Each cancels the CI and phone runs
+ before it. A ten-second local `cargo fmt --check && python3 scripts/locales.py --check` would
+ have stopped three of the four breaks.
+- **Phones:** the last phone success is `1d04d147` (10-23 04:45Z). It covers #507's Skia patch,
+ the open ladder, Back and long-press: arm64 16 KB check ok, API 30 and 35 smoke ok, iOS ok. Two
+ commits have never finished a phone build: `0c86f62` (mid-export fallback) and the spike.
+- **For a release:** a tag cut now fails at the CI-in-release gate. That is the gate working,
+ but nobody on `main` has a green commit newer than 10-23 03:11Z to tag.
+
+## 3. Review of the phone and CI commits
+
+### High
+
+**P1. Android's safe start leaves a real phone with no monitor and no export** (`7df9f8f`, #507;
+R from code, not observed). `safe_start` sets `CONCAT_GPU=software` on Android (`lib.rs:2656-2662`).
+The Skia patch reads that as "OpenGL ES for the window" (`patches/i-slint-renderer-skia/lib.rs:210`),
+which is what was intended. But the compositor reads the same variable as `software_only`
+(`concat-render/src/gpu.rs:828-835`), which means `force_fallback_adapter: true` on Vulkan and then
+on every backend. On Android the monitor has no window device (`platform.rs:671-685` returns
+`None`). Previews draw through `headless()` → `WgpuCompositor::new()`
+(`concat-export/src/lib.rs:1643-1666`), and exports through `best_compositor()`. A real phone has
+no CPU adapter: SwiftShader is an emulator thing, and the GLES adapter on Mali, Adreno or PowerVR
+is not typed Cpu. So both get `None`, previews log "nothing to draw on", exports fail with "this
+device's GPU could not be used to draw with", and the cards are left undrawn.
+
+- **It sticks for the whole build.** The GPU switch is hidden on Android
+ (`settings.rs:858`, `gpu_switchable: cfg!(any(windows, target_os = "linux"))`).
+- **What trips it:** any death in the first 15 s. That includes a swipe from Recents, a Force
+ stop, or a low-memory kill while the picker is open.
+- **CI cannot see it.** The emulator does have a SwiftShader CPU adapter.
+- **It may not fix #507 anyway.** #507 has no logcat (the reporter was asked 10-23), so the
+ Vulkan-crash diagnosis is a guess.
+
+Fix: on Android, have safe mode mean `Backends::GL` on hardware for the compositor too (the spike's
+`render.rs` already falls back that way), or take Android out of `safe_start` until there is a log.
+Show the switch on Android as well.
+
+**P2. Windows trust is now an outage, not a warning** (#533, V). SAC refuses unsigned DLLs once
+their reputation lapses. Every DLL beside `concat.exe` (BtbN FFmpeg, sherpa-onnx) is checked on
+its own, and SignPath Foundation will not sign upstream binaries (memory, `signpath.md`). Options:
+
+- **Microsoft Store (MSIX).** Registration is free for individuals. The `jub0t.Concat` identity
+ is still in repo variables (V). Store packages are signed by Microsoft, which covers every file
+ in the package (R; check SAC behaviour on an MSIX install).
+- **Azure Artifact (Trusted) Signing.** Paid. It signs any binary you ship, so the FFmpeg DLLs too (U on eligibility).
+- **Build FFmpeg in our own CI.** This is #533's option 1.
+
+Also, #533 says turning SAC off cannot be undone. Reports say the April 2026 update KB5083769
+made it re-enableable ([Topedia](https://blog-en.topedia.com/2026/04/smart-app-control-in-windows-11-can-now-be-re-enabled-without-reinstalling/),
+[Pureinfotech](https://pureinfotech.com/enable-smart-app-control-windows-11/)) (U). Check before
+telling users either way.
+
+**P3. Main is red, and the phone signal comes only from quiet hours** (§2). Still H4.
+
+### Medium
+
+**M1. Phone Back skips the phone editor's own sheet.** `app.slint:505-548` closes the desk
+sheets, then on Back with "nothing on top" runs `close-project`. The phone shell's sheet (Media,
+Text, Effects, Filters, every clip page) is a private `sheet` property of `PhoneEditor`
+(`phone/editor.slint:343`) that the handler never sees. So the most common Back on a phone, "put
+this panel away", saves and drops you on the launch screen (R). CapCut's Back closes the panel.
+
+**M2. The 0.2.8 notes still carry two false lines** ("Q and E cut", "Intel Macs open on macOS
+12 again"). #532 is a 0.2.8 launch death on macOS 13. Edit the body the way 0.2.7's was.
+
+**M3. The dependabot cargo group can never land.** It is one group for every pattern
+(`.github/dependabot.yml`), so majors and patches travel together:
+
+- Majors: ureq 2 → 3, cpal 0.16 → 0.18 (Android audio), jni-sys 0.4, pollster 1.0, dirs 7
+ (app-data paths), sha2 0.11, getrandom 0.4.
+- Patches stuck behind them: rustls, tokio, serde_json, and the sherpa-onnx crate 1.13.8 against
+ the 1.13.7 prebuilts in `deps.lock`.
+
+Three PRs in a row have been red on every job. Group by `update-types` (patch and minor
+together), and take majors one at a time with a review.
+
+**M4. The Compose spike sits in every CI job.** `concat-mobile` is a workspace member
+(`members = ["crates/*"]`). It is a cdylib over concat-host (gpu), export and media, plus uniffi
+0.32.2: 24 new lock entries, most of them bindgen-only. So clippy, check and nextest build and
+link it on Linux, macOS and Windows. It has 2 tests, and it already turned `main`'s fmt red. It
+ships in nothing: the shipped APK is `concat-android` via cargo-apk, and no workflow names
+`android-native` or `concat-mobile` (V, grep). Take it out of the workspace or out of
+`default-members` until it is a product decision. The cost per job was not measured (no builds
+here).
+
+Two smaller points on the spike:
+- Its Gradle wrapper has no `distributionSha256Sum`.
+- A binary `gradle-wrapper.jar` is committed with no validation step.
+
+**M5. The website goes stale at every release.** Nothing in `release.yml` triggers the website
+deploy. Add a `repository_dispatch` on release published, or a daily scheduled deploy.
+
+**M6. The community backlog grows faster than replies.**
+- 191 open issues (55 since 20 Oct). 92 of 150 outside issues (61%) have no maintainer reply;
+ 74 have no comment at all.
+- Five crash issues are unanswered: #294 iOS launch (10-09, oldest), #313, #372, #418, #423.
+- 7 of 9 contributor PRs have no review. #157 has waited 35 days. #513, #514, #529, #530 and #460
+ show only GitGuardian, meaning their CI waits on owner approval.
+- Of the 15 newest Discussions, 9 have no reply.
+
+### Low
+
+- **MediaCodec ladder (`2fb84cd`, #491).** Sensible order: as asked, then plain, then
+ `c2.android.*` / `OMX.google.*` by `codec_name`. Problems:
+ - The plain attempt clamps bitrate to 16 or 32 Mb/s and declares 30 fps for 60 fps input with
+ no word to the user. A frame-rate-sized rate control may overshoot.
+ - A 10-bit export still tries `c2.android.hevc.encoder`, which is 8-bit only.
+ - Tests check only the ladder table, and nothing has run on a device.
+- **Mid-export fallback (`0c86f62`).** The structure is sound: one retry, the stalled encoder
+ dropped first, the monitor's MediaCodec readers released at export start, and the stage named
+ for the sheet and for MCP. It restarts from frame 0, so a long phone export takes twice as long.
+ It has never finished a phone CI build.
+- **Long-press (`b467155`).** A two-finger pinch that starts on a clip and holds still for 400 ms
+ lifts it. There is no haptic at pick-up. The edge of a selected clip still trims on first
+ movement, with no slop.
+- **Gap closing (`c631c45`).** On a phone the magnet shows as On by default (`prefs.magnetic()`),
+ but trims read the raw field (`studio.rs:5110, 5242, 5826`) and do not ripple. Deletes off the
+ main lane do not close either. Settings says "deleting and trimming close the gap". One toggle
+ shows two states over three behaviours.
+- **#525 and #526** ("won't install") were answered well, with the exact size and SHA-256.
+
+## 4. Release and reach (V, `gh release view v0.2.8 --json assets`, 10-24)
+
+| Platform | Assets | Downloads | Share |
+|---|---|--:|--:|
+| Windows | setup.exe 24,212 + msi 3,030 | 27,242 | 44.5% |
+| Android | arm64 apk | 21,377 | 34.9% |
+| iOS | ipa | 4,472 | 7.3% |
+| Linux | deb 1,410+201, AppImage 1,229+646, rpm 322+106, pkg 254+72 | 4,240 | 6.9% |
+| macOS | arm64 3,095, x86_64 770 | 3,865 | 6.3% |
+| **Binaries** | | **61,196** | |
+
+- `manifest.json` was fetched 1,667 times and `SHA256SUMS` 1,137 times, so the all-asset total is
+ 64.0k.
+- Updater reach is about 2.7% of binary downloads, against 1% for 0.2.7.
+- The repo has 11,181 stars, 1,369 forks and 80 watchers.
+- Open issues by label: feature request 90, bug 77, ui-ux 38, windows 30, testing-requested 29,
+ crash 25, important 25, timeline 25, video 24, effects 21, android 20, priority 20, audio 18,
+ gpu 17, improvement 17, captions 15, AI 14, distribution 13, export 12, needs-info 11,
+ editor-api 11, linux 9, ios 8. None are unlabelled.
+
+## 5. Scores (0-10)
+
+| Aspect | 20 Oct | Now | One line of evidence |
+|---|:-:|:-:|---|
+| CI and release | 5 | **4** | The checks job and 16 KB gate are real, but 4 of 69 `main` CI runs are green, 58 of 65 phone runs were cancelled, `main` has been red about 29 h under four breaks, dependabot cargo has been red 3× and releases still pass unsigned |
+| Distribution and trust | 3 | **3** | Site current and vulnerability reporting on; but SAC now blocks installs (#533) on the 44.5% platform, there is no store, no Android registration, and the update manifest is unsigned |
+| Documentation (user-facing) | 5 | **5** | The site install guide is honest about SmartScreen/SAC/xattr; the README still ticks iOS ✅ and "4K scrubs smoothly", and the 0.2.8 notes keep two false lines that #532 disproves |
+| Design philosophy | 6 | **6** | Long-press, Back to projects, main-lane gap closing and a native Compose spike are platform-first; safe start ports a desktop "software" switch to a platform with no software adapter, and Back ignores the phone's own sheet |
+| User likeability on phones | — | **4** | What phones run is 0.2.8: a 4 KB library, Back quits, a tap moves clips; the fixes are 118 commits deep and unreleased, #507 is blank on a Redmi, #294 iOS is unanswered at 15 days |
+
+## 6. Before the next release
+
+1. Get `main` green: fmt, the `clipInspector.*` locale keys, the three clippy lints. Then let one
+ full CI run and one phone run finish on the commit to be tagged. Add a local pre-push check
+ for fmt and locales.
+2. Fix P1: on Android, the compositor uses hardware GLES in safe mode, and the GPU switch is
+ shown. Or take Android out of `safe_start`.
+3. Fix M1: Back closes the phone editor's sheet before it closes the project.
+4. Correct the 0.2.8 notes. In the next notes, name the macOS floor honestly, or make
+ `check-min-os.sh --enforce` pass.
+5. Windows: pick #533's path and start the Microsoft Store submission (the identity is
+ reserved). Say plainly in the README, site and notes that SAC blocks the current build.
+6. Wire the website deploy to the release.
+7. Register `app.concat.editor` for Android developer verification while the key is stable.
+8. Split the dependabot cargo group, and take `concat-mobile` out of the default workspace build.
diff --git a/audit/2026-10-24/window.md b/audit/2026-10-24/window.md
new file mode 100644
index 00000000..ac2b43bc
--- /dev/null
+++ b/audit/2026-10-24/window.md
@@ -0,0 +1,122 @@
+## Window crate (src/crates/concat), main 2606aff1, 2026-10-24
+
+Method: I read the code at `main` (2606aff1) and `git show` for every commit since c41865e that touches `src/crates/concat` (49 of the 102). I ran `scripts/locales.py --check` and `--same` (python only). I did not build, run cargo or open a device, so runtime behaviour is **reasoned** unless a line says otherwise. Slint behaviour was checked against `i-slint-core-1.18.1` in the cargo registry.
+
+Coverage: the 24 Oct UI commits, the 20-22 Oct fixes for H1/H2/H3, keys, copy/paste and prefs, and PRs #420, #421, #445, #446 and #498 were read in full. The four perf(media) commits, the SVG import and #485 were read for their window-side edits only. The engine reviewer found W-H1 on their own (engine.md E-H1).
+
+### Numbers (measured)
+| | 20 Oct (c41865e) | 24 Oct (2606aff1) | Δ |
+|---|---|---|---|
+| Rust lines (src/) | 33 216 | **34 619** | +1 403 |
+| Slint lines (ui/) | 34 957 | **35 929** | +972 |
+| studio.rs | 12 790 | **13 180** | +390 |
+| `impl Studio` | 9 887 lines, 248 methods | **10 262 lines (2316-12578), 263 methods** | +375, +15 |
+| `fn start` (lib.rs) | 2 242 | **2 342** (197-2538) | +100 |
+| lib.rs | 2 726 | 2 917 | +191 |
+| prefs.rs | 519 | 760 | +241 (atomic save, salvage, tests) |
+| Largest fns | publish_lanes 396, publish_chrome 353, key_editor_rows 330, stage_dragged 288, clip_set 264 | publish_lanes **405**, publish_chrome 353, key_editor_rows 330, stage_dragged 288, clip_set 267 | none split |
+| Largest .slint | media-pane 2 821, clip-inspector 2 642, start 2 574, lanes 1 926 | **clip-inspector 2 890**, media-pane 2 870, start 2 574, lanes 2 041 | clip-inspector +248, lanes +115 |
+| Non-test unwrap/expect | 4 | 4 | host.rs:174, studio.rs:2998, :10685, panes/start.rs:236 |
+| `unsafe` lines | 34 | 34 | |
+| `#[test]` | 155 in 28 files | **171 in 30 files** | +16 |
+| `accessible-*` props in ui/ | 19 in 3 files | 19 in 3 files | none on any control added this period |
+| Locales | 25 files | **26 (az)**, 908 keys, --check exit 0 | 985 values still English |
+
+New modules: speed_ramp.rs (331), section-tabs.slint (131), speed-ramp.slint (213). Nothing moved out of `Studio` or `start()`.
+
+### Status of 20 Oct findings
+| Finding | Status | Evidence (current code) |
+|---|---|---|
+| **H1** Safe start makes any early death permanent | **Partly** | Software now lasts one build, not for good: `prefs.rs:302-321` `settle_start`, keyed by `machine::build()`. `--smoke` no longer arms the mark (`lib.rs:214-216`). The notice shows on every launch it holds (`lib.rs:2664-2666`). Still open: any death of any cause in 15 s trips it (no crash signature). Errors returned before the loop still drop the mark through `Tripwire::drop` (`jobs.rs:169-172`; `select_backend(...)?` lib.rs:223-288, `Host::start` :290-296). Concurrent windows still share one mark, and the 15 s timer deletes it by path (`lib.rs:2496-2501`). New: on Android it now breaks export (W-H1) |
+| **H2** Android Back quits; nothing saved on the way out | **Partly** | Back takes Escape's road for window sheets (`app.slint:512-541`). With nothing on top it closes the project (`:547-553`). The loop's end saves on Android (`lib.rs:2521-2533`). Not covered: the phone editor's own sheets and the voiceover panel (W-H2). Slint still forwards no Pause or Stop, so a background kill loses up to the 1.5 s autosave |
+| **H3** Phone timeline moves clips on a tap | **Partly (body fixed)** | A press under a finger is nothing until it is a tap, a swipe past 8 px, or a 400 ms hold (`lanes.slint:585-600, 630-644, 669-686`). The floor's slop is 8 px (`:1282-1287`). A pinch's first finger no longer selects. Still: the selected clip's edge trims at once on the 4 px desk threshold (`lanes.slint:592`, `:657-658`). Targets are unchanged (`Theme.h` 28 px, `theme.slint:148`) |
+| Destructive chords while typing | **Fixed** (for the keymap) | `lib.rs:1712` drops every action but Save when `typing()` (`:2544-2551`) or `sheet_open()` (`studio.rs:9195-9201`). Ctrl+W, Ctrl+O and Ctrl+Q are KeyBindings (`app.slint:476-495`) and still fire from a field |
+| No guard while a sheet is open | **Partly** | `sheet_open` covers Export, Settings, Project, Relink and Confirm. Captions, Speech, Voiceover, the start sheet and the phone's sheets are not in it |
+| Capture `reserved()` unreachable; alias holes; letters dead on non-Latin layouts | **Unchanged** | keymap.rs changed only its copyright line; the KeyBindings above still match first |
+| Copy/paste/duplicate drop clip fields (N2) | **Fixed** | `match_command` (`studio.rs:2241-2251`) sends `MatchClip`, which copies the whole clip but its place (`concat-project/.../properties.rs:330-360`). Titles, shapes and now layers go through it too (`studio.rs:8340-8385`). A test covers mask, stabilize, opacity and transition (`concat-project/src/lib.rs`, `a_matched_clip_takes_everything_but_its_place`) |
+| Paste after switching projects (N1) | **Partly** | The clipboard carries the media item, and `held_here` maps it by path or adds it to the bin (`studio.rs:8288-8326`). `Enhanced { original, copy }` still holds the old project's media ids (W-M5) |
+| Prefs non-atomic, any parse failure resets all, world-readable (N4) | **Fixed** | Temp file then rename (`prefs.rs:352-365`). Written 0600 on Unix (`:391-401`). One bad value costs only itself (`salvage`, `:369-389`), and the bad file is set aside (`:337`). Residuals are Low (below) |
+| Interface scale not applied at launch | **Fixed** | `set_ui_scale` runs before the window (`lib.rs:203-211`). `apply_ui_scale` retries 40 × 50 ms until winit has a window (`platform.rs:66-75`) |
+| Toast not inset on phones (M3/N6) | **Unchanged** | `toast.slint:88`, 16 px off the window's bottom |
+| Select/SegmentedControl/Switch assign their own bindings | **Unchanged, new instance** | `select.slint:140`, `segmented-control.slint:47`, `controls.slint:563-565`. The new effect-card switch is one (W-M1) |
+| Export codec picker without `disabled` (M2) | **Unchanged** | `dialogs/export.slint:326-334` passes no `disabled` |
+| Duplicate is two undo steps (M4) | **Unchanged, wider** | `studio.rs:8183` `place_copy(..., None)`. Titles, shapes and layers are now two steps too, where titles were one (`:8381-8384`) |
+| Captions cancel/restart collide (L8); English error matching (N7) | **Unchanged** | `panes/captions.rs:229` `error.contains("cancel")` |
+| `--smoke` stops before the window (#17) | **Unchanged** | exits at `lib.rs:2454-2473`, before `app.run()` (:2509) |
+| Unknown argument refuses to start (#25) | **Unchanged** | `launch.rs:75-79` |
+| Event-loop failure exits with no dialog | **Unchanged** | `lib.rs:2513-2519` |
+| Quit during an export kills it without asking | **Partly** | `stop_exports` cancels cleanly and tells API callers (#431, `lib.rs:2586-2620`). It still does not ask |
+| RTL; accessibility names | **Unchanged** | No layout direction anywhere. The 19 `accessible-*` properties are the 20 Oct ones |
+| False SAFETY comments (L1, 12th) | **Unchanged + one more** | `lib.rs:2657` still cites "the scale above", which sets no env. The new `platform.rs:451-454` says "single-threaded" right after `Gpu::acquire()` (:436) has opened a device (lavapipe starts threads) |
+| English literals in notices (11th) | **Unchanged** | `studio.rs:4238`, `:8719-8720`, despite 72ba60a2 "every string the interface draws goes through the lookup" |
+
+Tally: 4 fixed, 7 partly, 11 unchanged (2 of them wider).
+
+### New findings
+**HIGH**
+
+- **W-H1. On Android, one early death means every export fails until the next build.** 7df9f8f2 runs safe start on Android (`lib.rs:2643-2645`). It sets `CONCAT_GPU=software` process-wide (`:2656-2662`), and the comment says that means "OpenGL ES in place of Vulkan (the Skia patch reads it), not the CPU". `concat-render/src/gpu.rs:828-833` reads the same variable as *software adapters only* for every compositor. Android lends the export no device (`platform.rs:681-684` returns `None`), so the export calls `best_compositor` → `WgpuCompositor::new()` (`concat-export/src/lib.rs:1573-1574, 1693`). With `force_fallback_adapter` there is no adapter on a real phone, so the export ends with "no GPU or software renderer". Cards go the same way (`concat-host/src/cards.rs:411-413`). Android has no GPU switch to undo it (`panes/settings.rs:858`). An early death does not need a bad driver: a swipe from Recents in the first 15 s, or the system killing the app behind the file picker, leaves the mark. CI's emulator runs SwiftShader, a CPU adapter, so CI would pass. *Reasoned; corroborated independently by engine.md E-H1, which adds that the monitor draws nothing.* Fix: on Android, give the Skia patch its own variable, or have `own_api_first` ignore `CONCAT_GPU` there.
+- **W-H2. Android Back on the phone editor's own sheets throws you out of the editor.** The phone's Media, Audio, Text, Effects, Filters, Mask, Canvas, Speed and other sheets live in `PhoneEditor`'s private `property <string> sheet` (`phone/editor.slint:343`). Nothing in `phone/` handles a key. The window's handler cannot see that state (`app.slint:512-541`), so Back falls through to `app-menu-selected("close-project")` (`:547-550`). Scenario: on Android, open Mask on a clip and press Back. The project saves and closes, and you are on the launch screen. The voiceover panel is missing from the list too: Back during a take closes the project, and the take is kept on disk but not laid on the timeline (`studio.rs:8813-8815`). 0aa1a502's message says "Back closes the sheet on top", which is untrue for the phone's commonest sheets. Fix: a `PhoneEditor.back() -> bool` that closes its sheet, called first; add `Editor.voiceover-open`.
+
+**MEDIUM**
+
+- **W-M1. The effect card's switch goes stale after Undo.** `Switch { checked: item.on; toggled => root.toggle(index) }` (`clip-inspector.slint:363-366`). The Switch writes `root.checked = !root.checked` itself (`controls.slint:563-565`), which ends the binding, and the rows are kept and `sync`ed in place (`studio.rs:803-818`). Scenario: switch an effect off, then Ctrl+Z. The effect is on and its knobs are bright (`opacity: item.on`), but the switch still shows off. The next click turns it "on" visually and off in the document.
+- **W-M2. Fold state belongs to the row, not the effect.** `property <bool> open: true` per repeated row (`clip-inspector.slint:331`), on a model updated in place. Fold the first effect and press its ↓: the effect moves down open, and the effect that took its place shows folded. Selecting another clip keeps the folds by index.
+- **W-M3. A sound layer made from a Clean effect cannot be seen, tuned or taken off.** The audio shelf carries `concat.denoise`, `concat.loudness` and `concat.enhance-voice` (category Clean). A sound layer's Audio tab is Layer + AudioEffects (`config-pane.slint:314`), and AudioEffects hides those three ids (`:798`) because a sound clip's Sound section has switches for them. A layer has no Sound section, so drag Denoise onto the lanes and the layer's only link appears nowhere. A Clean tile clicked with a sound layer selected disappears the same way.
+- **W-M4. The Effects tab appears on any clip whose colour was touched.** b5131917 shows Effects only "on a clip that already carries a chain (an older project, a template, the API)" (`config-pane.slint:251-255`). But the first Adjust knob inserts `concat.adjust` at the head of `video_effects` (`studio.rs:4755-4763`), and Basic › Mode › Chroma key adds `concat.chroma-key` (`clip-inspector.slint:1230-1238`). Both make `has-chain` true, mid-drag for the knob. The new tab then holds an "Adjust" card whose knobs are hidden (`studio.rs:1879-1883`) and whose trash removes the grade. On an image, the header flips from the plain heading with the clip's name to a segmented control. Also, desktop and phone now disagree: the phone still applies effects to the clip (`phone/editor.slint:1229-1230`).
+- **W-M5. A clip pasted into another project switches to the wrong file when Enhance is toggled.** `MatchClip` keeps `enhanced` (`properties.rs:330-360`). `held_here` remaps only `media_id` (`studio.rs:8288-8326`). `SetClipEnhanced` then sets `media_id` to `enhanced.original` or `.copy` (`concat-project/.../clips.rs:498-513`), which are the source project's ids. Fix: remap or drop `enhanced` in `held_here`.
+- **W-M6. Esc in the font picker's search no longer closes it, and the next key goes to the timeline.** SearchField now accepts Escape and calls `Focus.release()` (`controls.slint:763-770`), which focuses the window's `keys` scope (`app.slint:379-381`). Slint closes the top popup only on an *unhandled* Escape (`i-slint-core-1.18.1/window.rs:1372-1395`). The font list is a `PopupWindow` whose search is focused on open (`font-picker.slint:41-46, 98-101`). After Esc it stays open, the keyboard sits on the editor behind it, and `typing()` and `sheet_open()` are both false, so Q, W or Delete edit the timeline under an open popup. Enter has the same effect via `accepted`.
+- **W-M7. Every window focus stats every media file on the UI thread.** PR #421: `check_missing_media` → `missing_media()` on each focus-in (`lib.rs:256-264`, `studio.rs:8681-8689`). On a Windows project with footage on a share that has gone away, each Alt-Tab back waits on SMB timeouts. *Reasoned.*
+
+**LOW**
+- Keyboard reach lost in today's UI. The eye was an `IconButton` with a FocusScope and a name; the `Switch` has neither. The audio library's shelves were a SegmentedControl with arrow keys; the pills (`clip-inspector.slint:565-636`) and the card fold (`:351-354`) are TouchAreas. Bypass, fold and shelf choice now need a pointer.
+- Stale copy. `clipInspector.chainRunsTopBottom` still says "The eye bypasses one link" in all 26 locales (shown at `clip-inspector.slint:311`). `configPane.noEffectClipYet` still says "Pick one from the Library's Effects page", which now lays a layer instead. `Param.rail`'s doc says "not for a list of knobs" (`controls.slint:345-347`), and every effect knob now uses it.
+- Dead code and stale comments. `set_phone` still sets groups to -1 "a desk's strip starts on the first" (`studio.rs:9255-9262`), now the default for all. `library[1].category = "All"` duplicates group -1 (`:2455-2457`).
+- The audio pills do not wrap or scroll: about 330 px of pills against a 240 px seat minimum (`dock.rs:9`). A narrow inspector clips the last shelves. *Reasoned.*
+- A sound layer selected on the phone gets no clip tools: `phone/editor.slint:350-360` has no `sound-layer`. The phone cannot make one, but a desk project can bring one.
+- `typing()` reads `slint::private_unstable_api` (`lib.rs:2544-2551`), so a Slint bump can break the keyboard guard silently.
+- Prefs residuals. The `.json.unreadable` copy keeps the old file's mode through `fs::copy`, and it holds the API token (`prefs.rs:337`). A rename refused on Windows (AV or another reader) drops the save without a log.
+- `start.safeModePhone` (new, #507) is English in most locales. The localized "Untitled project" (`start.rs:207-225`) treats an English-named untitled project as typed after a language switch.
+- Doc drift outside the crate, from 5296d98c: `concat-host/src/cards.rs:506-509` now stacks the old "Every picture package has a card…" doc on `draw_every_card`.
+
+### Commits since c41865e, short verdicts
+| Commit | Verdict |
+|---|---|
+| f26ebf64 per-build safe start (#476) | Right shape; still trips on any death; see H1 |
+| 7df9f8f2 Android safe start (#507) | **Breaks Android export after one early death** (W-H1) |
+| 3f6db386 software adapter → Skia CPU + own device (#503, #480) | Sound idea; SAFETY wording false (above) |
+| 0aa1a502 Android Back + save on Destroy | Half: window sheets yes, phone sheets and voiceover no (W-H2) |
+| b467155c hold to move | Good; edges still 4 px |
+| e2be568a keys out of fields/sheets | Good; KeyBindings and four sheets not covered |
+| 1d04d147 MatchClip, paste by file | Good design (whole-clip copy, never on the API); `enhanced` not remapped (W-M5); duplicate still two steps |
+| 4b1b2b05 prefs atomic, 0600, salvage | Good, with tests |
+| 816a8f0e scale at startup (#462) | Fixed |
+| c495a8ea / #498 ⌘Q saves (#437) | Correct; a failed save there is only logged, which is unavoidable at terminate |
+| c631c459 phone delete ripples main lane (#497) | Read; matches phone editors; desk unchanged |
+| b3054fae field keeps caret in view (#506) | Read; no defect found |
+| 87fd6781 / #421 relink on focus (#409) | Works; UI-thread stats per focus (W-M7) |
+| 80561bd2 / #420 speed curve | Read speed_ramp.rs and the Studio edits; echo-then-commit like other knobs; no defect found |
+| 77f9457d / #445 early return republish | Correct |
+| 4ffa3212 / #446 quit cancels export | Correct; bounded at 10 s |
+| 0acb767b title Effects tab (#484) | Superseded two commits later by "only with a chain" |
+| bb30386b SectionTabs | Bar and index are clamped (`config-pane.slint:666-668`); arrows wrap; FocusScope present; no accessible role (none anywhere) |
+| 22639548 effect cards, audio pills | Looks better; W-M1, W-M2, keyboard reach lost |
+| 8dddc279 / b5131917 effects as layers, tab only with a chain | Product call is defensible (CapCut does both); execution leaks (W-M4); phone and desk now differ |
+| 0a45c5a5 SearchField Enter/Esc | W-M6 |
+| 6466b9f7 All rows, default group -1 | Correct (no `groups[group]` lookups in ui/); dead phone defaults left |
+| 599157e7 Preview chip, end_audition | Works (ends on clip press, deselect, Esc, cross, project close); locale keys added in all 26 |
+| 5296d98c square cards | Fine; edition bump redraws caches once |
+| 2606aff1 sound layers | Window side consistent (lanes, inspector, copy); W-M3 |
+
+### Scores (window crate)
+| Aspect | 20 Oct | 24 Oct | Evidence |
+|---|---|---|---|
+| Code quality | 6 | 6 | `MatchClip` and atomic prefs remove two bug classes. Today's cards reintroduce a self-assigning control (W-M1) and row-indexed state (W-M2), and an env var means two things in two crates (W-H1) |
+| Design architecture | 6 | 6 | A whole-clip copy command that never crosses the API is the right structure. `Studio` still absorbs everything: 263 methods, 10.3k-line impl, `start()` 2 342 lines |
+| Maintainability | 3 | 3 | +1.4k Rust and +1k Slint in 4 days with no split. publish_lanes is now 405 lines and clip-inspector.slint 2 890. Copy drift shows up in the same commit that changes the UI |
+| User likeability (UI/UX) | 6 | 6 | Real gains: hold-to-move, Back on sheets, safe start that heals, Preview chip, All shelves, square cards. Against them: Android Back ejects from phone sheets and an Android early death kills export, the Effects tab appears unasked, Esc misbehaves in the font picker, and keyboard reach was lost |
+
+### Before a release
+1. W-H1. Android export after a safe start. One line of scope (`cfg(not(target_os = "android"))` on the `CONCAT_GPU` read in `own_api_first`, or a separate variable).
+2. W-H2. Back closes the phone editor's sheet and the voiceover panel before it closes the project.
+3. W-M1 and W-M6 are small and user-visible. Bind the card switch one way with `changed`, or take the Switch's self-assign out. Let Esc through when the field is inside a popup.
diff --git a/audit/AUDIT-2026-10-24.md b/audit/AUDIT-2026-10-24.md
new file mode 100644
index 00000000..0500afbc
--- /dev/null
+++ b/audit/AUDIT-2026-10-24.md
@@ -0,0 +1,200 @@
+# Concat audit, 24 October 2026
+
+Code at `3e651f4` on `main`, 102 commits past the
+[20 October audit](AUDIT-2026-10-20.md) (`c41865e`). v0.2.8 is still the
+latest release: 117 commits on `main` have reached nobody.
+
+Four reviews feed this report, each read-only and each told to verify
+every claim in the code rather than trust commit messages, issues or the
+last audit:
+
+| Report | Covers |
+|---|---|
+| [`window.md`](2026-10-24/window.md) | The `concat` crate: shell, inspector, library, phone shell, keymap, prefs |
+| [`engine.md`](2026-10-24/engine.md) | concat-media, concat-render, concat-export, concat-core, concat-effects, host playback and posters |
+| [`host-api-security.md`](2026-10-24/host-api-security.md) | concat-host, server, API, MCP, project model, speech, security |
+| [`phones-ci-distribution.md`](2026-10-24/phones-ci-distribution.md) | Android, iOS, the Compose spike, CI and release, signing, website, GitHub |
+
+"Verified" means traced in code, in a shipped artifact, in CI logs or in a
+test that ran. Nothing was run on a phone or a Windows machine for this
+audit, so every on-device behaviour is reasoned, not observed.
+
+## 1. Where Concat stands
+
+**Reach keeps growing; trust does not.** v0.2.8 has 63 930 downloads in
+four days: Windows 27 268 (setup 24 238, MSI 3 030), Android 21 397, iOS
+4 472, macOS 3 095, Linux 2 639. 11 180 stars, 1 369 forks. Windows and
+Android are 76% of it. Every one of those installs is unsigned, or
+ad-hoc signed, and a new report (#533) shows the cost: Windows 11's Smart
+App Control now refuses Concat's unsigned FFmpeg DLLs outright on a
+machine where it ran a day earlier.
+
+**The engine and the edit got a great deal better in four days.** Seven
+format fixes (camera MTS/M2TS, MPEG-PS, HDV/DV pixel aspect, millisecond
+timestamps in MKV/WebM/FLV/WMV, variable-rate screen recordings, damaged
+packets, one time origin per file); captions timed by voice detection
+instead of 1-7 s early; posters in about 0.1-0.2 s instead of up to 17 s;
+a hardware encoder that stalls mid-export now finishes on the CPU;
+effects and sound effects are layers on the timeline; copy and paste
+carry every clip field; preferences are written atomically and privately.
+
+**The shell keeps absorbing it.** `studio.rs` is 13 180 lines (12 739 on
+20 October), `start()` 2 341 (2 240), the export's `lib.rs` 4 358
+(3 770). No file was split. The Compose spike showed why it matters: a
+native phone shell would have to re-implement the editing rules that live
+in `studio.rs`, because no UI-agnostic controller exists.
+
+**`main` gave no signal again.** CI on `main` was red from 23 October
+04:16 to 24 October (rustfmt drift, a locale key, three clippy lints),
+and almost every other run was cancelled by a newer push. Fixed in
+`e854f82`; the habit that caused it - pushing without the seconds-long
+local gates - is the finding.
+
+## 2. Scores
+
+| Aspect | 15 Oct | 20 Oct | Today | Why |
+|---|:-:|:-:|:-:|---|
+| Code quality | 6 | 6 | 7 ↑ | Whole-clip copy (`MatchClip`), atomic private prefs, engine fixes each with a regression test; against today's cards bringing back a self-assigning control and row-keyed state |
+| Design architecture | 7 | 7 | 6 ↓ | Effects and sound effects as timeline layers, one mix path for preview and export; `Studio` at 263 methods and 13 180 lines, with no controller a native shell could share |
+| Design philosophy | 5 | 6 | 6 | Layers over menus, a Preview chip, All on every shelf, Back that closes sheets; against a desktop safe mode ported to Android that broke its exports (fixed today) |
+| Maintainability | 5 | 4 | 3 ↓ | +1.4k Rust and +1k Slint in the window crate in four days, export `lib.rs` 4 358 lines, `clip-inspector.slint` 2 890; nothing split |
+| User likeability | 6 | 6 | 6 | Large wins on `main` - captions on time, posters at once, formats, layers - none of it released; users still run 0.2.8 |
+| Performance | 5 | 6 | 7 ↑ | Posters 12-17 s → 0.1-0.2 s, phone filmstrips 11-14 s → 0.15 s, one reader per file in playback; scrub mode (#472) still unbuilt |
+| Scalability | 5 | 5 | 5 | Fixed 512 MB export pool on any device, three MediaCodec seats first-come; per-project ids fixed for clips but not Enhance or backgrounds |
+| Security | 5 | 5 | 5 | Prefs 0600; an updater trust regression (46af73c) was found and fixed today (41ca63a); the manifest is still unsigned and the API findings untouched |
+| Testing | 5 | 6 | 7 ↑ | 1 080 tests (+~100), real-media fixtures for formats, export A/V and sound-layer tests; still no Slint UI tests and no device run |
+| Documentation | 6 | 5 | 5 | Honest commit history and notices; the 0.2.8 notes still carry two false lines and README ticks run ahead |
+| CI and release | 4 | 5 | 4 ↓ | 4 green CI runs on `main` in 69 since 20 Oct, 58 of 65 phone runs cancelled; green again from `e854f82` |
+| Distribution and trust | — | 3 | 2 ↓ | Smart App Control now blocks the unsigned build outright (#533); no store, no Android registration, unsigned manifest |
+
+## 3. What landed since 20 October
+
+Grouped by what a user sees. Every item is on `main` and in no release.
+
+| Area | What changed | Evidence |
+|---|---|---|
+| Captions | Timed by Silero VAD, each caption drawn in to its speech: the reporter's 0/7/18/36/47 s became 3.07/14.08/25.09/36.10/47.11 s against speech at 3.1-47.1 s; upstream whisper.cpp gap fix (#528) | `2922614`, `ea22531`; synthetic test only |
+| Formats (#457) | TS/PS seeks, HDV/DV pixel aspect, millisecond stamps, VFR screen recordings, damaged packets, one time origin per file; 52-file matrix against ffmpeg's own decode | `d02b756`-`93c18b7` |
+| Speed | Posters on every worker ahead of filmstrips, one keyframe in software; keyframe filmstrips | `7b2d30e`-`7dc91a9` |
+| Export | Frame count from the timeline's own grid (#531); chained transitions keep their lanes (#527); a stalled phone encoder reruns once in software | `053a4fb`, `d935ca4`, `0c86f62` |
+| Playback | A transition within one file keeps a reader per side (#505), without opening three for plain playback | `6c1c181`, `2b6eb23` |
+| Editing | Effects and sound effects dragged onto the lanes as layers; picture effects only as layers; a Preview chip; All on every shelf; square bright cards; inspector section tabs, effect cards, search fields; title Effects tab | `8dddc27`-`6466b9f`, `2606aff`, `599157e` |
+| Data | Copy, paste and duplicate carry every clip field; prefs atomic, 0600, salvaged per value | `1d04d14`, `4b1b2b0` |
+| Platform | macOS: no libavdevice (#532, #342); Windows on WARP: CPU window, own monitor device (#503, #480); Android: Back, long-press move, gap-closing delete, MediaCodec open ladder (#491) | `af90722`, `3f6db38`, `0aa1a50`, `b467155`, `c631c45`, `2fb84cd` |
+| Spike | Kotlin + Compose shell over the engine through UniFFI, on the emulator: GPU preview with no readback, 26-30 fps, 29 MB APK against 80 MB | `77ce50b`; emulator only |
+
+Fixed today from this audit's own findings: the Android safe start that
+left the compositor with no adapter (`5bef0c2`), the updater trusting a
+re-signing proxy's manifest (`41ca63a`), Back leaving the project from a
+phone sheet, the effect switch after undo, invisible Clean sound layers,
+the Effects tab on graded clips, Esc in the font picker (`3e651f4`), MCP's
+`preview_frame` hint, the VAD model's shared `.part`, and a red `main`
+(`e854f82`).
+
+## 4. The 20 October High findings
+
+| 20 Oct | Then | Now |
+|---|---|---|
+| H1 Safe start permanent | High | Partly: one build, `--smoke` clean, notice each launch; any death in the first 15 s still trips it. Its Android port broke every export (fixed today, `5bef0c2`) |
+| H2 Android Back quits | High | Partly: window sheets and phone sheets close (`0aa1a50`, `3e651f4`), saves on Destroy; Back during a voiceover take still leaves the take off the timeline |
+| H3 Phone timeline moves on tap | High | Partly: 400 ms hold to move, 8 px swipe slop; a selected clip's edge trims at 4 px, targets still 28 px |
+| H4 CI on `main` no signal | High | Unchanged in effect: 4 green in 69; fixed today, but the cause was pushing without the local gates |
+| H5 No trust chain | High | Worse: #533 Smart App Control refusals; the updater regression is fixed; manifest unsigned, no store, no Android registration |
+| H6 macOS floor | High | Partly: libavdevice gone (#532 proved the crash on 13); the version check still only warns |
+| H7 Export recovery on hybrid laptops | High | Unchanged (`concat-export/src/lib.rs:1612-1624`, `gpu.rs:960-963`) |
+| H8 Captions early | High | Fixed in code; measured on synthetic speech, not yet on user audio |
+
+## 5. Open findings, High
+
+- **O1. Windows installs are refused (#533, distribution).** Smart App
+ Control blocks the unsigned BtbN FFmpeg DLLs, on machines where 0.2.7
+ ran. Windows is 44% of downloads. SignPath Foundation will not sign
+ upstream binaries; the ways out are the Microsoft Store (the
+ `jub0t.Concat` identity is reserved) or FFmpeg and sherpa-onnx built in
+ our own CI and signed with the rest.
+- **O2. The update manifest is unsigned (H5).** It is now fetched over the
+ shipped roots alone, which closes today's regression, but anyone who can
+ publish a release can still ship any installer. A compiled-in key and a
+ signature on the manifest close it.
+- **O3. Export recovery takes the road that crashes hybrid laptops (H7).**
+ Unchanged since 15 October.
+- **O4. 117 commits reach nobody.** The macOS 12/13 launch crash (#532),
+ the WARP crashes (#503, #509, #523), the Android encoder stall (#491)
+ and every fix above are on `main` only, and every report about them
+ will keep coming until a release carries them.
+
+## 6. Open findings, Medium
+
+Full evidence in the area reports.
+
+- **Engine.** The playback reader's slack is counted in source frames, so
+ 60-240 fps clips, speed ramps and 1080i still open two or three readers
+ (E-M1). Sound seeks never check where they landed, so a TS/PS file
+ trimmed from the front can play silence at its head (E-M2). The phone's
+ software rerun goes to `c2.android.*` encoders, which likely cap below 4K
+ and at 8 bits, so a stalled 4K or 10-bit export renders twice and fails
+ (E-M3, unverified). No deinterlacer: 1080i shows combing. TS scrubbing
+ ~40 ms a frame against ~15 for MP4.
+- **Data.** A paste into another project carries the source's media ids
+ in Enhance and in image backgrounds (W-M5, M6). 0.2.8 opening a project
+ with a sound layer drops it, and its next autosave deletes it (M8): a
+ project format version that older builds refuse would prevent it.
+- **Window.** A card's fold is remembered by row, so it moves with a
+ reorder (W-M2). Every return to the window checks every media file on
+ the UI thread, a freeze on Windows network drives (W-M7). Ctrl+W, Ctrl+O
+ and Ctrl+Q still fire from a text field. Keyboard access to the new
+ switch, pills and fold, and the "eye" in the chain's help text, in all 26
+ languages, are Low.
+- **Server and agents.** Unchanged since 20 October: one shared auth
+ throttle bucket, agent-opened projects locked until the server ends, the
+ project lock opening on any I/O error, `concat --mcp` never run in CI, no
+ size caps on downloads, phone FFmpeg and sherpa pulled without digests,
+ the model mirror 404.
+- **Phones.** The toast sits under the navigation bar; no soft-keyboard
+ inset, RTL or accessibility names. The Compose spike is a workspace
+ member, so every CI job builds and links it.
+- **Speech.** The VAD step cannot be cancelled and shows no progress.
+
+## 7. GitHub and community
+
+- 191 open issues, 10 open PRs. A backlog sweep on 24 October found about
+ 60 of the open issues already fixed (36 in a release, 25 on `main` only),
+ duplicated (15) or confirmed fixed by their reporters (#223, #297, #324,
+ #389); 27 are real open bugs and 79 feature requests. Closing them is
+ blocked by this session's permission settings and waits on the owner.
+- Every issue filed through 24 October is labelled and answered, including
+ the eight filed on 23-24 October. Labels: bug 77, ui-ux 38, windows 30,
+ testing-requested 29, crash 25, important 25, android 20.
+- Merged since 20 October after review: #360, #425, #426, #505 (with a
+ fix for the regression found in review), #527, #528, #531. #529 is merged
+ on a working branch with styled text ranges built on it. Changes
+ requested or waiting: #383 (licence, due to close ~29 October), #513/#514
+ (audio-only export), #530, #508, #460, #358, #157.
+- Flags: an outside offer by email to publish Concat on Google Play under
+ the sender's own developer account and signing key (Capotis Apps) was
+ advised against: it would hand them the listing and the update channel.
+ #520's claims about the API were checked and mostly wrong (transitions,
+ audio clips and stills are in the document model). No prompt injection
+ found.
+
+## 8. What to do, in order
+
+1. **Release from a green `main`.** Let one full CI run and one phone run
+ finish on the tagged commit. The crash fixes above reach nobody before.
+2. **Choose the #533 road** - Microsoft Store submission, or FFmpeg and
+ sherpa-onnx built and signed in CI - and say plainly in the release
+ notes and on #533 that Smart App Control blocks the current build.
+3. **Sign the update manifest** with a key compiled into the app.
+4. **Register as an Android developer** for `app.concat.editor` before
+ Google's verification rollout.
+5. **Extract the editor controller from `studio.rs`** into a crate with no
+ Slint in it: it is the maintainability score's only real fix, it makes
+ the inspector testable without a window, and the native phone shell
+ needs it. The spike's estimate: 4-6 weeks.
+6. **The Mediums:** E-M2 (check sound seeks), E-M1 (slack in output
+ frames), M6 (remap Enhance and backgrounds on paste), a project format
+ version (M8), W-M2, W-M7.
+7. **Close the 55 issues the sweep found done or duplicated** (a script is
+ ready) and correct the two false lines in the 0.2.8 notes.
+8. **Before every push:** `cargo fmt`, the locale check and clippy on the
+ crates touched. Seconds, and they keep `main` a signal.