The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.
Rust ffmpegfree-video-editormodel-context-protocolnon-linear-editoroffline-firstopen-source-video-editorself-hostedvideo-editingvideo-editing-automationvideo-editing-softwarevideo-editing-toolsvideo-editorvoice-cloningyoutube-automationyoutube-shorts
docs(audit): the 20 October audit - where Concat stands after v0.2.8
Jareer998bf301b9cd
diff --git a/audit/2026-10-20/host-server-api-cli.md b/audit/2026-10-20/host-server-api-cli.mdnew file mode 100644index 00000000..412361b8--- /dev/null+++ b/audit/2026-10-20/host-server-api-cli.md@@ -0,0 +1,175 @@+## host, server, API, CLI, MCP: read at ce64dae (2026-10-20)++Read-only. Nothing built or run. "Verified" means I read it at the line given. "Reasoned" means inferred from code I read. "Unverified" means I could not check it.+Since 10-15: 27 commits touch concat-host, 7 concat-api, 3 concat-server, 4 concat-mcp (new crate), 1 concat-cli. +8 586 / -499 lines across the five crates (`git diff --stat ab21457 HEAD`).+The server step that slipped in both earlier audits has now landed (363800f), and the new MCP crate is a thin adapter of good quality. What is still broken: how claims and the lock interact with agent sessions, a throttle that any local process can turn against the user, and MCP annotations that misstate what tools do.++### Numbers (`#[test]` count, `unsafe ` count, `.unwrap()` count in src/)+| crate | lines | tests | unsafe | unwrap |+|---|---|---|---|---|+| concat-host | 17 541 (+~5 600) | 125 | 46 | 34 |+| concat-server | 2 513 (+1 120) | 27 | 0 | 0 |+| concat-api | 2 971 (+619) | 26 | 0 | 0 |+| concat-mcp | 3 041 (new) | 11 (3 e2e) | 0 | 0 |+| concat-cli | 520 (+3) | 3 | 0 | 0 |+Biggest files: concat-api lib.rs 1 782, playback.rs 1 347, tools.rs 1 133, media.rs 1 109, titles.rs 878, projects.rs 867, updates.rs 862.++---++### Re-verification of every 10-15 item++**10-15 new findings**+| # | Item | Status | Evidence |+|---|---|---|---|+| 1 (M6) | pre-auth fd exhaustion | **FIXED** | registry entry dropped with the connection: `concat-server/src/json.rs:158-169, 292-301`. A seat is taken only after auth (`:275`). Pending callers capped at 32 with a 5 s whole-line deadline (`:241-249, 261, 342-370`). Accept backs off on errors (`:206-224`). Tests `:578, :598, :618` |+| 2 (M13) | crash handler takes locks / allocates | UNCHANGED | `crash.rs` has no commits since 10-15 |+| 3 (M14) | Android second Vulkan device | UNCHANGED | `preview.rs` off_thread untouched (diff is only `forget_plan` `:375-384`). The cards still build a compositor every round (`cards.rs:410`). d9f7e08 only skips software adapters |+| 4 | api lib.rs:7 "an MCP server" | MOOT | MCP now exists (`concat-mcp`) |+| 5 | `connections()` counts subscribers | **FIXED** | counts authenticated seats: `server lib.rs:339-342` |+| 6 (L11) | `ConsolidateClips` adopts any file | UNCHANGED, **wider reach** | `concat-project/src/commands/mod.rs:630`, `clips.rs:670`. An agent can now reach it through MCP `edit_apply` (`tools.rs:241-247`) |+| 7 (L12) | weak tokens, no throttle, 0.0.0.0 | PARTLY | `check_token` (`token.rs:32-46`). The window refuses a weak token (`settings.rs:637`). Throttle added (`throttle.rs`). Off-loopback warning exists (`settings.rs:876`, `is_loopback :1053`). But the estimator passes guessable tokens (new L2), and the throttle is a new DoS (new M1) |+| 8 | two crash codes can never arrive | UNCHANGED | crash.rs untouched |+| 9 | merge bakes un-analysed masks | unverified (window code, not in scope) | — |+| 10/11 | mono output loses right channel; RT-thread frees | UNCHANGED | `playback.rs` no commits |+| 12 (M16) | README promises gRPC + CLI | **FIXED** | `README.md:68` "A build from source adds gRPC and a command-line tool". Releases still build `-p concat` only (`build-app.yml:550`) |+| 13 (L15) | download size unbounded / Windows single rename | size UNCHANGED (`models.rs:224-269`), now reused by sounds (`sounds.rs:432`). Rename **FIXED** | 5 retries with backoff on Windows (`projects.rs:357-372, 388-392`) |+| 14 | merged renders never collected | unverified (window) | — |++**Older items carried by 10-15**+| Item | Status | Evidence |+|---|---|---|+| 12Oct#5 24 held stills FIFO | unchanged (pool not re-read) | `preview.rs:167-170` hold_still as before |+| 12Oct#6 shared save scratch | **FIXED** | per-call `create_new` temporary with pid and nonce (`projects.rs:304-345`). Ticket ordering under a per-folder mutex (`:181-298`). Commits e75af9b, 9197340. The cross-process case is now covered by the lock. The doc at `projects.rs:265-271` still says "a project folder has none [lock] today" |+| 12Oct#30 512 MiB live-title ceiling | UNCHANGED | `titles.rs:98` |+| 12Oct#32a `gpu` capability never appears | UNCHANGED | `Api::with_dirs` uses `Monitor::new()` (`api lib.rs:296`), which sets `gpu: None` (`preview.rs:92-100`), so `has_gpu()` is always false (`:156-165`). `lib.rs:506` never fires. The test pins its absence (`lib.rs:1262`) |+| 12Oct#32b stop waits on open gRPC Events | UNCHANGED (reasoned) | The Events sender lives in the hub subscriber list (`grpc.rs:75-77`). Graceful shutdown waits for the stream (`:170-173`). `close_doors` joins the listeners (`lib.rs:404-406`) before `hub.close()` (`:420`). Only gRPC embedders and tests are affected; the window has no gRPC |+| 12Oct#32c CLI render in place, no MAX_SIDE | UNCHANGED | `concat-cli/src/main.rs:419-431` |+| 12Oct#33 dirs.rs "downloaded models" | **WORSE** | `dirs.rs:32` still says "Large state: downloaded models", but `data` now also holds irreplaceable user content: the audio library and cloned voices (`dirs.rs:89-100`, `sounds.rs`, `voices.rs:75-77`). On Windows that is non-roaming %LOCALAPPDATA% |+| 11Oct#21 docs | PARTLY | api lib.rs:7 now true. `concat-host/Cargo.toml:37` still claims a CPU compositor |+| 11Oct#6 stop hangs on wildcard bind (Windows) | **FIXED** | `loopback()` (`server lib.rs:89-100, 393-395`), test `:515` |+| 11Oct#7 refused thread kills accept loop | **FIXED** | `json.rs:252-256`; errors back off `:218-221` |+| 11Oct#10 API create unchecked | UNCHANGED | `api lib.rs:516-533` calls `projects::create` with no `checked_size`/`checked_rate`. A rate_den of 0 is written, `open` then refuses it (`projects.rs:453`), and the folder is left behind |+| 11Oct#15 server edges | PARTLY | gRPC slow subscriber silently dropped: unchanged (`grpc.rs:75-77`). A poisoned subscriber list silences events: unchanged (`hub.rs:54-57`). CLI `api` dies on non-UTF-8: unchanged (`cli main.rs:175-176`). key_of doc on `impl Drop`: unchanged (`api lib.rs:1064-1066`) |+| 11Oct#16 stop blocks window; jobs die silently | stop **FIXED** (`stop_detached` `server lib.rs:369-379`, used at `settings.rs:624`). Jobs: PARTLY | `reap` books a vanished job as failed (`api lib.rs:897-907`). Under `panic = "abort"` (`src/Cargo.toml:127`) a panic still ends the process. `playback.rs:683, 693` unwraps unchanged |+| 11Oct#17 reads unconfined / caller strings to FFmpeg | UNCHANGED, now documented | `api lib.rs:360-365`. `preset` passes through unvalidated (`:708-711`) |+| 11Oct#23 host lows | UNCHANGED, one **WORSE** | The silent data-folder move (`dirs.rs:124-136`) now also moves `audio`/`voices` user content (`:111-116`). `updates.rs` has no commits. `templates.rs:357` poster as preview.jpg |+| 11Oct#25 WAV header overflow | **FIXED** | `record.rs:621-626` clamps to `u32::MAX - 36`; header `:590-606` |+| 11Oct#29 `--socket` deletes whatever file | UNCHANGED | `server lib.rs:272` |+| S1 unsigned update chain, quarantine strip | UNCHANGED | `updates.rs` has no commits since 10-15. Still this area's worst standing risk |+| S2 API root = whole home | UNCHANGED, wider | `server lib.rs:131-137, 206-210`. MCP headless uses the same default (`concat-mcp/src/backend.rs:185-188`) |+| S3 | **FIXED** | see #1 |+| S4 Clear cache deletes reversed/enhanced media | unverified | `reverse.rs` untouched; `enhance.rs` changed and was not re-read for this |+| S6 embedded API has its own monitor/GPU | UNCHANGED (reasoned) | `api lib.rs:296`. It also shares the process-wide decode pool with the window (`preview.rs:94`), so an agent's previews evict the user's frames |+| S8 HDR proxy name | UNCHANGED | `proxy.rs` untouched |+| S9 composites on hub thread; slow subscribers | hub UNCHANGED (`api lib.rs:840-868`). JSON PARTLY | A full outbox (256) now silently unsubscribes the caller while its connection stays (`json.rs:316, 447-455`, `hub.rs:58-66`). A writer stalled more than 10 s is hung up on (`json.rs:63, 430-437`) |+| S13 per-sample clip scan, S14 silent missing font | UNCHANGED | files untouched |+| S16 connections count | **FIXED** | — |+| S17 Unix socket perms from umask | UNCHANGED | `server lib.rs:273` (no chmod) |++Tally, about 40 items: 11 fixed, 6 partly, 3 worse, about 17 unchanged, the rest unverified or moot. Every 11-Oct and 12-Oct server-step item that was in scope landed in 363800f, and each landed with a test.++---++### New findings++**HIGH: none new.** S1 (the unsigned update chain, unchanged) remains the area's top risk.++**MEDIUM**++**M1. The auth throttle turns loopback into a shared lockout: any local process, or a web page, can lock the user's own agent and scripts out of the window's API indefinitely.** Verified.+- Every TCP caller is keyed by IP alone (`json.rs:93-97`), so all local callers are `Peer::Ip(127.0.0.1)`. Every Unix-socket caller is one `Peer::Local` (`:114-116`).+- Any first line that is not a correct `auth` counts as a failure (`json.rs:267-271`), and that includes `GET / HTTP/1.1`.+- After 3 failures the address is refused before its token is read (`json.rs:228-240`), with waits doubling to 60 s (`throttle.rs:110-116`). The test pins that even the right token is refused (`json.rs:650-659`).+- Failure: a browser tab runs `fetch("http://127.0.0.1:7420")` four times, or any unprivileged local process sends one garbage line a minute. `concat --mcp` is then refused permanently. In `auto` mode it silently falls back to headless (`backend.rs:81-86`, stderr only), which is the path the window mode exists to avoid.+- Second vector: 32 silent sockets renewed every 5 s fill `MAX_PENDING`, and everyone gets `busy` (`json.rs:241-249`). Silent connections are not even throttled (`:262-266`).+- The throttle buys nothing for a minted 128-bit token; the module says so itself (`throttle.rs:6-9`).+- Fix: skip the throttle when the token was minted. Count only well-formed auth attempts. Cap pending connections per peer.++**M2. Projects an agent opens stay claimed and locked until the server stops or the process exits. The window then refuses them with a message that offers no way out, and the agent's unsaved edits are discarded with no warning.** Verified.+- Claims are released only by `project.close` or when the `Api` is dropped (`api lib.rs:579-592, 1066-1073`).+- A connection that ends does nothing to sessions (`json.rs:332-335`).+- The Window backend's `finish` is the no-op default (`backend.rs:47`; `Window` does not override it, `:143-167`).+- A headless `concat --mcp` keeps its `concat.lock` for the whole agent session, because clients keep MCP servers alive.+- Failure: the agent builds "Promo" through the window's server and the session ends. The user clicks Promo in Recents and gets "This project is open through the Remote API" (`en.json:721`, `studio.rs:8169-8171`). The only way to free it is to switch Automation off, and that drops the agent's unsaved edits without a word (the `Api` drop does not save).+- Every keystroke in the host, port or token field restarts the server (`settings.rs:356-372` → `apply_server` → `stop_detached`). Each restart discards every agent session and cancels its export.+- Fix: track sessions per connection and close (or save-prompt) them when it ends. Have `Window::finish` close what it opened. Let the window take over an API-held project with a prompt. Apply settings changes on Enter, not on every keystroke.++**M3. MCP annotations misstate what tools write, and writes reach anywhere under home.** Verified.+- `preview_frame` is `readOnlyHint: true`, `destructiveHint: false` (`tools.rs:445-453`). With `output` it calls `File::create` on any path under the roots, with no extension check and no check for an existing file (`api lib.rs:811-813, 1148-1163`).+- `export_start` claims `destructiveHint: false` (`tools.rs:413-419`) but overwrites an existing `output` (`api lib.rs:665`, muxer guessed from the extension).+- `project_open` is unconfined, and `project_save` then writes `concat.json`, plus the lock and caches, into that folder wherever it is (`api lib.rs:360-365, 600-602`).+- Failure: a client that auto-approves read-only tools (behaviour varies by client, unverified per client) runs an agent that is prompt-injected through title text or media names. Those strings flow into tool output via `summary.rs`. It calls `preview_frame{output:"~/.ssh/config"}` and the file is replaced by a PNG with no prompt.+- `docs/MCP.md` presents "only under your home folder" as a safety property; home is where everything that matters lives.+- Fix: correct the hints. Require `.png` and a new file, or an explicit `overwrite`. Refuse to export over an existing file unless asked. Consider defaulting the roots to Videos/Movies plus the open project folders.++**M4. A lock that fails with an I/O error lets the open go ahead with no guard.** Path verified; Windows behaviour reasoned.+- `OpenProjects::claim` maps `LockError::Io` to "no lock, open anyway" (`api lib.rs:183-189`).+- `acquire` turns `PermissionDenied` into `Io` after 4×50 ms (`lock.rs:97-106`).+- On Windows a lock file held without sharing by an AV scanner, OneDrive or Search, or left delete-pending on volumes without POSIX delete semantics (FAT/exFAT, SMB, older builds; unverified which builds), gives exactly that error.+- Windows `same_file` is only `path.exists()` (`lock.rs:257-263`), so a lock taken on a delete-pending file passes the check. Every later opener then gets `PermissionDenied` → `Io` → an unguarded open. This is two processes saving one folder, the thing 68ceffe set out to prevent, on the priority platform.+- Network file systems with no locks degrade silently too.+- Fix: refuse, or open read-only, when the folder is writable but cannot be locked. Retry `PermissionDenied` for longer.++**M5. `concat --mcp` as shipped is never run by CI, and its panic safety net is dead code.** Verified.+- The shipped profile is `panic = "abort"` (`src/Cargo.toml:127`). That makes `catch_unwind` in `protocol.rs:426-429` ("a panic in a tool is that call's failure") and `hub.rs:91-104` (inside the window's server) no-ops.+- Any engine panic kills the agent session and its running export. Inside the window it kills the window and the user's unsaved work.+- The release binary is GUI-subsystem on Windows (`concat/src/main.rs:8`); its use as a stdio server is untested.+- The e2e test drives the dev `concat-mcp` binary (`tests/e2e.rs:29`). The workflows contain no `--mcp` (grep of `.github/workflows`).+- The discovery → `Window::find` path has no end-to-end test either: e2e sets `CONCAT_API_ADDRESS`/`TOKEN` (`e2e.rs:385-391`).+- Fix: a CI smoke test that runs `concat --mcp headless` from the app-profile build on Windows and macOS, sending `initialize`, `tools/list` and one edit.++**LOW**+- **L1.** The Window backend re-sends any request once after a connection error (`backend.rs:144-162`). If the server applied it and the reply was lost (server restart mid-call, or the 10 s write-timeout hang-up `json.rs:430-437`), a `clip_add` or `captions_add` lands twice. The hang-up is reachable: the client reads events only during a call (`client.rs:95-115`), so an idle agent during an export backs up the 256-line outbox. Retry only idempotent tools.+- **L2.** The token estimator (`token.rs:53-72`, floor 64) passes `12345678901234567890` (about 66 bits) and `abcdefghijklmnopqrst` (about 86). On a 0.0.0.0 bind, IPv6 address rotation evicts strikes once 4 096 peers are booked (`throttle.rs:76-89`). Any success from a shared IP wipes its strikes (`json.rs:273`, `grpc.rs:204`).+- **L3.** Inline previews are unbounded. `width`/`height` up to 8192 (`tools.rs:878-904` defaults only when both are absent) gives a ~256 MB RGBA frame, PNG-encoded and base64'd into one JSON line headed for the agent's context. Above the client's 256 MiB `MAX_REPLY` (`client.rs:21`) the line is cut, read as "not JSON-RPC", and retried through L1, so the frame renders again.+- **L4.** Recents are written non-atomically (`projects.rs:545-553`) by the window, the window's API and every headless MCP process (`api lib.rs:570-572`). A read during a write returns empty, and the next `remember` wipes the list. Agent projects also fill the user's Recents and home page.+- **L5.** `Api::cancel` cancels the shared export slot (`api lib.rs:786-789`), not its job's flag, contrary to `job.rs:13-16`'s own rationale. A small race can cancel the window's export.+- **L6.** Job names restart at `j1` for each `Api` (`api lib.rs:892-895`). After a server restart, an agent polling `j1` gets "no job" or someone else's job.+- **L7.** The Tripwire decides by pid liveness (`jobs.rs:157-165`), while `lock.rs:18-23` explains why pids answer wrongly. On Windows and Android a reused pid lets a crash-loop job restart itself.+- **L8.** No read timeout after auth (`client.rs:55-59`). With one request at a time (`protocol.rs:18-22`), a stuck hub or a 600 s `export_wait` blocks every other tool, `export_cancel` included.+- **L9.** Two windows: the second overwrites `automation.json`, and when its server stops it removes the listing (`discovery.rs:134-141`), so the first window's running server is no longer listed.+- **L10.** Sound downloads trust a digest from the same manifest as the URL, with no size cap (`sounds.rs:383-395, 432-447`; `models.rs:224-269`). Same class as S1.+- **L11.** Stale or misplaced docs:+ - `prefs.rs:230-231` says an empty token means "none"; it means a minted one.+ - `projects.rs:265-271` says there is no lock.+ - `settings.rs:964-969`: `new_token`'s doc sits on `LOG_LEVELS`.+ - `settings.rs:1044-1052`: `split_listen`'s doc is merged into `is_loopback`'s.+ - `api lib.rs:1064-1066`: key_of's doc sits on `impl Drop`.+ - `protocol.rs:426` is false under abort.+ - `concat-host/Cargo.toml:37` still claims a CPU compositor.+ - `dirs.rs:32` "downloaded models".+ - `is_exposed` (`server lib.rs:77-82`) is unused; the window duplicates it as `is_loopback`.+- **L12.** The MCP "modern" protocol `2026-07-28` with `server/discover` (`protocol.rs:36-41`) is past my knowledge: **unverified** against the spec.++### Commits since 10-15, area view+- **363800f server hardening.** Correct in shape: RAII seats and places, a registry that cleans itself up, a whole-line auth deadline, a bounded outbox with a write timeout, and stop cancelling before joining. Each guard has a test (`json.rs:545-778`). What it gets wrong is the throttle model (M1).+- **56ff055 discovery + client.** Owner-only 0600 write via staging and rename (`discovery.rs:76-110`); removal checks the pid. Fine. Windows relies on the folder's ACL (unverified).+- **db529b2 / f70332c / 30584f7 MCP.** A thin adapter, with schemas generated from the API types and round-trip tested (`tools.rs:1087-1116`). Cancellation and progress are handled properly. Annotations wrong (M3); shipped path untested (M5).+- **68ceffe / 46e04b6 project lock.** An OS lock rather than a pid file, takeover after a crash, an inode recheck against the release race (`lock.rs:113-118`), a real child-process test (`:300-345`). Holes: M4, and M2's lifetime problem.+- **e75af9b save ordering.** Fixes 12Oct#6 properly.+- **Host features** (sounds, voices, stabilize, enhance preview, tripwire, voiceover recorder): skimmed for path and security handling only. File names are sanitised (`sounds.rs:162-168, 224-252`; `voices.rs:89-99`) and the voice decode is bounded (`voices.rs:117-120`). Not audited for behaviour.++### Good+- The fd-leak fix is complete and tested.+- Stop is detached from the UI thread.+- The lock is designed properly.+- Saves are ordered, with their own temporaries and Windows retries.+- MCP schemas come from the types, so they cannot drift.+- Stdout carries protocol only: no `println!` in the engine crates outside tests (grep).+- The discovery file is owner-only.++### Scores (area; 10-15 is the mean of the host, server, API and CLI rows)+| Aspect | 10-20 | 10-15 | Why |+|---|---|---|---|+| Code quality | 8 | ~7.3 ↑ | The server rewrite is careful RAII with no unwraps. Dead `catch_unwind` and misplaced docs remain |+| Design architecture | 7 | ~7.3 → | One dispatcher with thin transports, MCP as an adapter: right. Claims are not tied to connections (M2), the lock fails open (M4), the embedded API has its own monitor and the shared pool |+| Design philosophy | 8 | ~6 ↑ | "Refuse rather than lose", an OS lock rather than pids, schemas from types. The Tripwire contradicts it (L7) |+| Maintainability | 7 | ~7.3 → | Clear modules, but `api lib.rs` is 1 782 lines and `tools.rs` 1 133; docs drift (L11) |+| User likeability | 6 | (Remote API 4) ↑ | MCP is real value with a one-line setup. Dead-end refusals (M2), Recents pollution (L4), settings edits that kill agent work |+| Performance | 6 | ~5.7 → | One hub thread composites synchronously; previews unbounded (L3); an agent shares the window's decode pool |+| Scalability | 5 | 5 → | One API thread, one export slot, 64 seats: fine for a local tool, nothing more |+| Security | 5 | ~4.5 ↑ | Pre-auth fd leak fixed, token floor, owner-only discovery. A new lockout DoS (M1), annotations that understate writes plus home-wide roots (M3), and S1 untouched |+| Testing | 7 | ~5.5 ↑ | 27 server tests, a cross-process lock test, MCP e2e headless and via server, schema round trips. The shipped `--mcp` path, the discovery path and Windows lock races are untested (M5, M4) |+| Documentation | 6 | 6 → | `docs/MCP.md` is clear and accurate on modes, but overstates the safety of "home". About ten stale or misplaced comments (L11) |++Top fixes, in order: M1 (no throttle on a minted token), M2 (claims tied to sessions), M3 (correct hints and refuse overwrites), M5 (CI smoke of `concat --mcp` on Windows), M4 (fail closed), then S1.diff --git a/audit/2026-10-20/media-render-export-core.md b/audit/2026-10-20/media-render-export-core.mdnew file mode 100644index 00000000..be698da2--- /dev/null+++ b/audit/2026-10-20/media-render-export-core.md@@ -0,0 +1,186 @@+## media, render, export, core (+ vision frames): HEAD ce64dae, 2026-10-20++Labels: **Verified** = I read it at the line. **Reasoned** = I followed the code path; nothing was run. **Suspected** = the reason is given. Nothing was built or run (read-only audit). wgpu-hal / wgpu-types 30.0.1 sources were read from `~/.cargo/registry`.++Since 2026-10-15 these crates have 26 commits: +6 296 / −252 lines over 32 files (`git diff --stat 11e1bd3 HEAD`). The ones that matter:+- 2969e02: Android decodes on MediaCodec.+- ac8a654: feed.rs.+- b18c2fe and a9b2df6: the FXC pin.+- d3c2dcd: GLES uniform alignment.+- b5e6c08: the safe-start hook in render.+- 89258aa: average bitrate.+- 4f69340 / 95d6ea0: the VFR rate.+- a72e8ec: `Rational::parse`.+- 16a3d13: masks.+- 03dac1f: backgrounds.+- 72cbfe5 / 4edd466: stabilization.+- 0f63fca: adjustment-layer test.+- 4451548: #448 test.+- ce64dae: transition knobs.++### 1. Previous findings re-verified++| ID | Status | Evidence at HEAD |+|---|---|---|+| 12Oct#4 FXC compile unbounded, repeated, Cancel deaf | **UNCHANGED, exposure WORSE** | `shader()` compiles synchronously with no deadline (`concat-render/src/gpu.rs:2615-2690`, `block_on(scope.pop())` :2681). Cancel is checked only at the top of a frame (`concat-export/src/lib.rs:1687`). Every compositor starts with an empty map (`gpu.rs:1387`), and `sibling()` makes a fresh one (:960-963), so every export recompiles. b18c2fe/a9b2df6 now force FXC on every Windows machine (`gpu.rs:861-862`, `concat/src/gpu.rs:69-70`). b5e6c08 sends crashed launches to WARP+FXC for good (see H1). |+| 12Oct#7 GLES 30 s wait is really ~2.15 s | **UNCHANGED, exposure WORSE** | wgpu-hal 30 clamps it: `timeout_ns.min(i32::MAX)` (`wgpu-hal-30.0.1/src/gles/fence.rs:162`, `device.rs:1757-1763`). `DEVICE_WAIT` still documents 30 s (`gpu.rs:288-296`). d3c2dcd makes effects and transitions actually run on GLES, so slow GLES frames now happen in practice (M3). |+| 12Oct#9 HEVC/AV1 export opens NVENC first | **UNCHANGED** | `concat-media/src/encode.rs:128-139`. The export still passes `hardware: true` (`concat-export/src/lib.rs:1648`). |+| 12Oct#10 a validation error counts as a lost GPU, then second device, then WARP | **UNCHANGED** | `gpu.rs:3543-3573`; `concat-export/src/lib.rs:1521-1550`. |+| 12Oct#11 `combine` unscoped, refused passes skipped silently, lent counter never moves | **UNCHANGED** | `combine` has no error scope (`gpu.rs:3584-3653`). Refused passes: `gpu.rs:2686, 2809`. `with_device` counter: `gpu.rs:1413`. The window device's handler only logs (`concat/src/gpu.rs:133`), so on Windows (the lent sibling) an uncaught error never fails an export frame. |+| 12Oct#31 8-bit phone fallback clips HDR | **UNCHANGED** | Warn only: `gpu.rs:903-911`. |+| 12Oct#33 DEVICE_WAIT doc and two read_back docs | **UNCHANGED** | `gpu.rs:288-296`. |+| 12Oct#5 a preview layer that fails on the export side is dropped silently | **UNCHANGED** | Failures are collected but only reported when *every* layer fails (`concat-export/src/lib.rs:2296-2309`). |+| 12Oct§7 encoder EAGAIN retry | **FIXED (as before)** | Still no phone export in CI. |+| 11Oct#8 uncaptured GPU errors | **PARTLY (as before)** | Remainder is 12Oct#11. |+| 11Oct#11 mux scratch names collide | **UNCHANGED** | `concat-export/src/lib.rs:1180-1182`: two exports of the same stem in the same folder clobber each other. |+| 11Oct#12 paced and pooled decode disagree past clip end | **UNCHANGED** | Uses the output frame's duration (`decode.rs:1183-1190`). |+| 11Oct#13 `pick_rate` VFR | **PARTLY FIXED** | 4f69340/95d6ea0 round an oversized average instead of falling to the 90 kHz grid (`probe.rs:241-245`, `concat-core/src/time.rs:361-386`). Still no VFR media fixture, only arithmetic tests. A VFR clip now reports e.g. 61.219 fps (unverified whether the timeline ever adopts it). |+| 11Oct#14 `MAX_SIDE` doc | **UNCHANGED** | `concat-core/src/frame.rs:27-31` still says every GPU meets 8192. GLES phones are 4096 (`gpu.rs:970-972`). |+| 11Oct#21 CPU fallback docs | **UNCHANGED** | `concat-export/src/lib.rs:19` ("CPU compositor as the always-correct fallback") against `gpu.rs:808` ("since the CPU compositor went"), while :782 and :1418 still promise CPU answers. `render()` returns black (`gpu.rs:3532`). |+| 11Oct#22 dead core code | **UNCHANGED** | `concat-core/src/timeline.rs:329` (`source_duration` is used only by tests) and :510-517 (`Project`, unused). |+| core `Neg`/`ceil` on `i64::MIN` | **UNCHANGED** | `time.rs:158-160, 251-258`. (`parse` itself was fixed in a72e8ec, verified.) |+| N1 Android never decodes on MediaCodec | **PARTLY FIXED** | 2969e02 opens `h264/hevc/vp9/av1_mediacodec` by name for 8-bit 4:2:0 (`hardware.rs:239-250`, `decode.rs:684-731`). The log now names the decoder only after a real open (`hardware.rs:427-434`). **Still CPU:** 10-bit HEVC (`hardware.rs:241`), which is the HDR phone footage that most needed the chip; keyframe-only readers; and the 4th reader onward. The real-file device test is `#[ignore]` (`decode.rs:1850`) and appears in no script or workflow (`grep` of `scripts/` and `.github/`: zero hits). New problems in this code: M1, L1. |+| N2 chained transitions climb into the next track's lane | **UNCHANGED, now worse in effect** | `concat-export/src/lib.rs:889` (`b.track = a_track + 1` with `a.track` already bumped). Still no test. A packaged span's `to_track` (lib.rs:808) then lands on track t+1's lane, and `transition_stages` forces every layer on that lane to opacity 1 (`lib.rs:2008-2014`). A caption on track t+1 is pulled into the transition at full opacity. |+| N3 HDR proxy opens NVENC in the background | **UNCHANGED** | `concat-host/src/proxy.rs:238` (`hardware: hdr`). The proxy decode is now software (:218), which is good. |+| N4/M7 audio mix holds the timeline's sound, Cancel deaf | **UNCHANGED** | `concat-export/src/lib.rs:1202-1210`, `audio.rs:371+`. |+| N5/M8 crossfade heard only in the file | **UNCHANGED** | `resolve_transitions` is still private (`lib.rs:621`). Playback uses `flattened_clips()` (`concat-host/src/session.rs:306`). |+| N6/M12 `encodable()` opens MediaCodec encoders on the UI thread | **UNCHANGED** | `encode.rs:162-180`; the doc admits "before the window shows". |+| N7/M9 FFmpeg ERROR/WARNING dropped | **UNCHANGED, impact WORSE** | `ffi.rs:101-105`. It now also hides the MediaCodec *decoder's* configure and dequeue reasons. Only "would not open … Invalid argument"-style summaries reach the log. |+| N8 `EffectClock` doesn't forward `lost`/`deliver_hdr` | **UNCHANGED** (latent) | `lib.rs:1905-1940`. |+| N9 decoders drop a packet on EAGAIN | **FIXED** (ac8a654) | `feed.rs:76-107` keeps and resends. Every decode loop goes through it: no `send_packet` call is left outside feed.rs. Fake-codec tests are in `feed.rs:176+`. |+| N10 ~45/80 MB allocated per 4K frame | **UNCHANGED** | `gpu.rs:3492-3510`, `encode.rs:1552-1559`. |+| N11 prefetch "single-threaded" doc | **UNCHANGED** | `prefetch.rs:241` against `decode.rs:826-837`. |+| N12 Android SDR proxy holds a MediaCodec encoder | **UNCHANGED** | Android's list is MediaCodec-only, whatever `hardware` says (`encode.rs:106-112`). |+| N13 volume doc drift | **UNCHANGED** | `animate.rs:334` `volume_at` still has no caller. |+| M14 second Vulkan device per export or card round | **UNCHANGED** | `concat-host/src/preview.rs:146-153`. |++**Tally (28):** 1 fixed (N9), 1 still fixed (§7), 3 partly (N1, 11Oct#13, 11Oct#8), 19 unchanged, 4 unchanged with worse exposure (12Oct#4, 12Oct#7, N2, N7).++### 2. New findings++#### HIGH++**H1. On a Windows machine whose last launch died while starting, every export runs on WARP+FXC permanently. Effect compiles then take minutes with Cancel deaf.** *Verified path; the timing comes from the repo's own note.*+- How it happens:+ - b5e6c08's safe start (`concat/src/lib.rs:2459-2474`) sets `CONCAT_GPU=software` for good whenever the previous process died within 15 s, for *any* reason: a driver, a panic=abort in project load, Task Manager, a power cut.+ - `own_api_first` honours it for the compositor too (`concat-render/src/gpu.rs:831-833`).+ - WARP compiles through FXC (b18c2fe pin, `gpu.rs:861-862`).+- The repo's own test note says that on WARP each package compile "ran past five minutes … every compile going through FXC … and never finished" (`gpu/tests.rs:35-40`). That is why those tests skip WARP (`heavy_gpu`, `tests.rs:41-48`).+- An export compiles every effect again (12Oct#4), synchronously, before the frame's cancel check (`concat-export/src/lib.rs:1687`).+- **Scenario:** a user whose GPU driver once crashed at start exports a clip with a Gaussian blur, glow or blurred background (03dac1f puts `concat.gaussian-blur` behind every "Blur" background, `lib.rs:1465-1470`). The progress bar sits at 0 for minutes and Cancel does nothing.+- The only way back is a Settings switch the user has no reason to find.+- **Fix:** keep the compositor on the GPU when only the window crashed, or at least probe the GPU again. Compile off-thread with a deadline. Cache pipelines across exports (`PipelineCache`, or keep the sibling's shader map).++**H2. On Windows, the export draws on a sibling of the window's device. Any GPU error in the window marks the export's frame failed, and recovery opens the second device that #202 says kills NVIDIA laptops.** *Partly verified, partly reasoned.*+- `sibling()` shares the `uncaught` counter (`gpu.rs:960-963`).+- `render()` counts any increment during the frame as a failed frame and kills the compositor (`gpu.rs:3543-3573`).+- For the lent (window) device the counter never moves (12Oct#11), so this is latent on that exact path.+- Separately, any readback timeout or OOM on the lent device sends the export to `recovered(1)` = `WgpuCompositor::new()` (`concat-export/src/lib.rs:1530-1531`): a fresh HighPerformance DX12 device. On hybrid laptops that is the dGPU while the window is on the iGPU, which is exactly the "second road through the driver" that `render_on`'s own doc (`lib.rs:1072-1078`) says ended in a crash (#202).+- **Scenario:** an 8 GB card runs out of memory mid-export (#223/#297 class). Instead of retrying on the window's device, the export opens NVIDIA directly and the app dies with nothing in the log.+- **Fix:** recover first on `Monitor::off_thread()` again (a new sibling of the window's device), then WARP. Never `new()` when the export was lent a device.++#### MEDIUM++**M1. MediaCodec seats are first-come and never reclaimed. Idle warm readers and background jobs hold all three, so the clip under the playhead decodes on the CPU for the whole life of its reader.** *Verified path; impact reasoned.*+- 3 seats process-wide (`hardware.rs:268`). Seats are taken at open (`decode.rs:690-697`) and released only when the reader drops or falls back.+- Warm readers are closed only after 10-15 s idle (`concat/src/lib.rs:2388-2403`, `pool.rs:599-622`). A software reader never moves onto the chip when a seat frees.+- `DecodeOptions::default()` is `HwPolicy::Preference` (`decode.rs:152-158`), so other jobs take seats too:+ - stabilization analysis (`concat-host/src/stabilize.rs:174`);+ - filmstrip `read_through` (`concat-host/src/media.rs:490`);+ - `still_at` one-frame grabs (`media.rs:824`), the "chip's worst case" the commit excluded only for `keyframes_only`;+ - cutout and brush jobs (`cutout.rs:279`, `brush.rs:206`);+ - the export's own decoders and its private pool (`concat-export/src/lib.rs:1684, 1774`).+- **Scenario:** a phone edit of short clips (the TikTok shape). After three cuts, every newly opened reader is software, so playback stutters exactly where the commit claims a win. During export, preview readers or a stabilization pass hold the seats, so the export decodes 4K on the CPU.+- **Fix:** give seats by priority (playhead and export over background), make background and one-frame readers `in_software()`, and let a reader take a freed seat at its next seek.++**M2. Any live adjustment layer turns every packaged transition into a dissolve, so today's knobs (ce64dae) silently do nothing under a timeline-wide grade.** *Verified.*+- `composite_treated` skips the shader when *any* treatment covers the instant (`concat-export/src/lib.rs:1958-1964`). The monitor does the same (`lib.rs:2174-2180`).+- The comment calls it "a rare pairing". A grade adjustment layer over the whole edit is the most common way adjustment layers are used (0f63fca just tested that path).+- `transitions.iter().find(..)` (`lib.rs:1960`, :2181) also runs only the *first* live span, so two transitions overlapping in time on different tracks give one shader and one dissolve.+- Only `transition_pass` and `combine` are tested with knobs (`gpu/tests.rs:838-864`, `lib.rs:3346-3354`). No test checks that knobs survive `composite_treated` or the monitor.+- This is the code-level candidate for #448 ("all transitions look the same, just a fade"), which 4451548 could not reproduce because its test bypasses this gate. Other candidates are a refused pipeline, which falls back silently (`gpu.rs:2804-2810`), and a GLES device before d3c2dcd.++**M3. A GLES phone's export dies when one frame takes over 2.15 s. That is now reachable because d3c2dcd lets effects run on GLES.** *Verified wgpu clamp; scenario reasoned.*+- The 30 s `DEVICE_WAIT` becomes `i32::MAX` ns (`wgpu-hal gles/fence.rs:162`).+- The first "loss" opens a new device. The second asks for `software()`, which an Android phone does not have, so the export fails: "the GPU device was lost part way through the export, and this device's GPU could not be used…" (`concat-export/src/lib.rs:1530-1537, 1496-1497`).+- **Scenario:** a Mali/Adreno GLES fallback phone exports 1080p with a blurred background plus an effect. Two frames over 2.15 s and the export is gone, on the platform with the most downloads.+- **Fix:** on GLES, poll in a loop up to `DEVICE_WAIT`, not one clamped wait.++**M4. Export bitrate moved from CRF to single-pass ABR at YouTube's *upload* rates, with x264 `veryfast`. That is a visible quality drop for the default export.** *Verified settings; the quality impact is the well-known behaviour of x264 1-pass ABR.*+- The default is now `RateMode::Average` with `preset "veryfast"` (`concat/src/panes/export.rs:441-452`). 1080p30 H.264 gets 8 Mb/s (`export.rs:336-353`).+- x264 runs `b/maxrate 1.5×/bufsize 2×` with no CRF (`encode.rs` `libx264 … if average`, 89258aa).+- Previously it was CRF 16, which the sheet's own estimate put at about 16 Mb/s at 1080p30 (`encode.rs` `implied_kbps` doc).+- Single-pass ABR is x264's least efficient mode: it starves the first seconds and busy scenes. These are master files that platforms re-encode, so the generation loss compounds.+- Defensible as a product choice for predictable size, but it should be CRF-capped VBV (`crf` + `maxrate`) or 2-pass, not bare ABR.++**M5. The export's private reader pool is a fixed 512 MB regardless of device memory.** *Verified.*+- `ReaderPool::with_defaults()` (`concat-export/src/lib.rs:1684`) gives 512 MB and 8 readers. It does not use `sized_for` (`pool.rs:589-591`). It sits on top of the app's pool (256 MB to 1 GB).+- It fills with any reversed or speed-curved clip, and with background stills (`lib.rs:1449-1452`). At 4K that is 15 frames (8-bit) or 7 (deep).+- **Scenario:** a 3-4 GB phone exports a reversed clip next to the open editor and gets OOM-killed. That compounds N4 (the audio mix), which is still unbounded.++**M6. A blurred background runs the clip's whole effect chain twice per frame.** *Verified.*+- `backdrop()` clones `self.effects` and appends the blur (`concat-render/src/plan.rs:463-471`).+- **Scenario:** a vertical export with a "Blur" background over a clip with a glow and a LUT costs twice the effect time per frame, plus a full-frame Gaussian. On WARP or GLES that doubles H1 and M3.+- **Fix:** blur the effected picture once, then reuse it for both.++#### LOW++- **L1. The MediaCodec first-picture probe proves a codec once per process on one file.** *Reasoned.*+ - `Verdicts` is keyed by decoder name only (`hardware.rs:361-381`), probed on whichever file came first (`decode.rs:1248-1281`).+ - A later stream that wedges the wrapper (another profile or level, or session pressure) spins inside FFmpeg with the reader's lock held. `STUCK_AFTER` (`feed.rs:68`) guards only EAGAIN rounds, not a wedge inside `send_packet`.+ - The probe also opens a session outside the seat budget, and a wedged probe leaks its thread, file and session for the life of the process.+ - The verdict mutex is held across the 2 s wait (`hardware.rs:369-381`), so up to 2 s per codec on first open (suspected acceptable).+- **L2. `next_frame` can spin without counting.** *Suspected; depends on wrapper behaviour.* When `send(End)` returns EOF it does not count toward `STUCK_AFTER` (`feed.rs:96`). A codec that answers EAGAIN to receive after draining would busy-loop forever. libavcodec's contract says it shouldn't.+- **L3. Transition knobs are neither clamped nor sanitised.** `Package::resolve` passes set values through (`concat-effects/src/catalogue.rs:655-671`). `Clip::tidy` touches only `duration` (`concat-project/src/model.rs:2032-2034`). The current shaders have fixed loop bounds (verified: blur-dissolve, whip-pan, zoom-blur), so this costs odd pictures, not hangs. A community transition with a knob-bounded loop is not GPU-trialled (`trial_at` builds an effect layer only, `gpu.rs:1808-1861`).+- **L4. The monitor falls back to three CPU round trips for every transition frame whenever any layer sits above the incoming lane** (`concat-export/src/lib.rs:2182-2189`): captions or titles, i.e. most edits. Playback stutter at cuts on weak GPUs (reasoned).+- **L5. A stabilized clip's transition handle can fall outside what was measured.** The analysis margin is 1.5 s (`concat-host/src/stabilize.rs:53-55`), while an overlap's head can be half of any length (`concat-export/src/lib.rs:874`). Before the first measured time the correction is `IDENTITY` (`concat-vision/src/stabilize/path.rs:95-107`). So a 4 s cross-fade into a stabilized clip shows its first 0.5 s unzoomed and unsteadied, then pops (suspected).+- **L6.** `WGPU_DX12_COMPILER=dxc` re-opens the PATH-DLL crash that #418 fixed (`gpu.rs:861-862`, `concat/src/gpu.rs:69-70`). Intentional, but undocumented outside a comment.++### 3. Commits worth noting (verified)++- **2969e02 (MediaCodec decode):** careful.+ - Packet timebase is set so timestamps round-trip exactly for every container clock coarser than 1 µs (`decode.rs` `pkt_timebase`).+ - Colour is read off the stream, not the wrapper (`decode.rs:741-765`).+ - Flush happens only after output (`decode.rs:934-937`); I checked that `next_source` returns only after output, so the "flush before first output loses CSD" case cannot arise.+ - The filtergraph is rebuilt from the frame's real format (`decode.rs:1032-1041`), so NV12 output is fine.+ - Gaps: M1, L1, and 10-bit staying on the CPU.+- **ac8a654:** the right fix, written once, with fake codecs.+- **b18c2fe/a9b2df6:** correct for #418. wgpu 30's default really is `Auto` (`wgpu-types-30.0.1/src/backend.rs:801-803`). Both instances are pinned. It does nothing for #297.+- **ce64dae:** the plumbing is correct (`flatten.rs:136`, `lib.rs:661, 810`), and the shader params buffer is rewritten every draw and submitted immediately (`gpu.rs:2893-2900`), so different knobs on the same key don't bleed. Weakened by M2 and L3.+- **d3c2dcd:** effect Frame is 32 bytes. The transition Frame is already 16 (`concat-effects/src/shader.rs:182-189`, `gpu.rs:2813`). Consistent.+- **16a3d13 (masks):** the WGSL matches core (`gpu.rs:87-169` against `concat-core/src/mask.rs:179-217`). The feather band is floored at a pixel, so `smoothstep` never gets `low == high`.++### 4. Numbers (HEAD; counts are grep counts, test files included)++| crate | LOC | tests | ignored | unsafe | unwrap/expect | panic (non-test) |+|---|---|---|---|---|---|---|+| media | 11 413 (+1 085) | 113 | 1 | 40 | 222 | 0 |+| render | 8 685 (+962) | 101 | 0 | 1 | 104 | 0 |+| export | 4 832 (+592) | 38 | 0 | 0 | 46 | 0 |+| core | 3 547 (+650) | 60 | 0 | 0 | 26 | 2 (arena unreachable) |+| vision | 5 498 (+2 738) | 59 | 1 | 0 | 31 | 0 |++Biggest files: export `lib.rs` 3 770 (+366), `gpu.rs` 3 695 (+134), `encode.rs` 2 553, `gpu/tests.rs` 2 361, `decode.rs` 2 032 (+275).++### 5. Scores (area; the 10-15 figure is the mean of that audit's per-crate scores)++| Aspect | 10-15 | Now | Why |+|---|---|---|---|+| Code quality | 7 | 7 | feed.rs and the seats/verdicts code are clean and tested with fakes; the god files keep growing. |+| Design architecture | 7.5 | 7 ↓ | Seats have no priority (M1); the shader-or-dissolve gate (M2); recovery opens the second device (H2); still no pipeline cache. |+| Design philosophy | 7 | 7 | Honest fallbacks and honest commit text (10-bit kept on the CPU, said why), but "rare pairing" and "30 s" are untrue. |+| Maintainability | 5.5 | 5 ↓ | export `lib.rs` 3 770 and `gpu.rs` 3 695; the stale docs from 11 and 12 Oct still stand. |+| User likeability | – (~6) | 6 | Knobs, masks, backgrounds, stabilization and phone 8-bit hardware decode are real wins; HDR phone footage is still CPU; under a grade the knobs do nothing; safe-start users get exports that hang. |+| Performance | 6 | 6 | Android 8-bit decode up a lot; offset by seat starvation, WARP+FXC, doubled effects under backgrounds, three readbacks per transition frame. |+| Scalability | 5 | 5 | Audio mix memory; fixed 512 MB export pool; recompile per export; three sessions. |+| Security | 7 | 7 | All unsafe carries SAFETY; the FXC pin removes a DLL-hijack crash; knobs unvalidated (L3). |+| Testing | 7 | 6.5 ↓ | Good unit fakes; but the MediaCodec device test is never run; no chained-transition test; no knob test through `composite_treated`; heavy GPU suites skip Windows/WARP, exactly where H1 lives. |+| Documentation | 5 | 5 | New code well commented; `DEVICE_WAIT`, `MAX_SIDE`, CPU-fallback and prefetch docs still false. |++### 6. Unverified or not checked++- MediaCodec behaviour on real phones: the only evidence is the commit text.+- Whether a VFR clip's 61.219 fps rate ever becomes a timeline rate.+- x264 ABR quality was not measured.+- FXC compile times on real (non-WARP) Windows GPUs (#297).+- concat-vision enhance (03c8fd6) was not read.diff --git a/audit/2026-10-20/product-distribution.md b/audit/2026-10-20/product-distribution.mdnew file mode 100644index 00000000..f2c87da8--- /dev/null+++ b/audit/2026-10-20/product-distribution.md@@ -0,0 +1,220 @@+# Product, community and distribution audit, 20 October 2026++Repo `jub0t/Concat` at `ce64dae`. v0.2.8 was published at 2026-10-20T08:18:58Z from tag `v0.2.8` = `79116bd` (2026-10-18). Its assets were uploaded at 2026-10-19T03:38Z. Sampled 08:40–10:50Z. Read-only: `gh` reads, `curl`, `apksigner` on downloaded APKs (stored in the scratchpad, not the repo). Issue and PR text was treated as data.++Legend: **V** = verified here; **U** = unverified; **R** = repeated from `reports/` and not re-checked.++## Bottom line++Demand is huge, and distribution throws much of it away. v0.2.7 drew 126k binary downloads in a week. The v0.2.8 launch is being undercut by Concat's own website. The "Start here" site still serves **v0.2.7 links** next to a `latest/` checksum file that now lists **0.2.8** names. In one 15-minute window, v0.2.7 gained as many downloads as v0.2.8 did (+68 vs +75).++Concat sits in no store or package manager it runs itself. Windows is unsigned, macOS is not notarised, iOS is ad-hoc, and the Android signing key changed between 0.2.5 and 0.2.6. Two-thirds of outside-filed open issues have no maintainer reply. Three issues marked `testing-requested` today have their fixes **only after the v0.2.8 tag**, so no reporter can test them. SECURITY.md points reporters at GitHub's private advisory form, but **private vulnerability reporting is disabled**.++The smaller rival Drift is already on Flathub, the Microsoft Store, a Homebrew tap and Google Play closed testing.++## 1. Releases and reach++### Downloads (V, `gh api repos/jub0t/Concat/releases --paginate`; manifest.json and SHA256SUMS excluded)++| Release | Published | Total | Windows | Android | iOS | macOS | Linux |+|---|---|--:|--:|--:|--:|--:|--:|+| v0.2.8 | 10-20 08:18Z | 584 → 659 at 10:50Z | 295 | 163 | 10 | 52 | 64 |+| v0.2.7 | 10-13 | 126,218 | 54,353 (43.1%) | 44,190 (35.0%) | 10,279 (8.1%) | 9,242 (7.3%) | 8,154 (6.5%) |+| v0.2.6 | 10-06 | 11,328 | 5,155 | 3,323 | 243 | 793 | 1,814 |+| v0.2.5 | 09-28 | 42,954 | 17,665 | 15,206 | 3,487 | 3,512 | 3,084 |+| v0.2.4 | 09-24 | 4,364 | 2,049 | 833 | 195 | 284 | 1,003 |+| v0.2.3 | 09-19 | 11,846 | 5,919 | 2,718 | 716 | 1,126 | 1,367 |+| v0.2.2 | 09-15 | 3,316 | 2,017 | 535 | 59 | 332 | 373 |+| v0.2.1 | 09-04 | 4,277 | 2,567 | 913 | 81 | 423 | 293 |+| **All** | | **≈204.9k** | | | | | |++- Per asset for v0.2.7: `windows-x86_64-setup.exe` 48,897, `android-arm64.apk` 44,190, `ios-arm64.ipa` 10,279, `macos-arm64.dmg` 7,498, `.msi` 5,456, `x86_64.AppImage` 2,546, `linux-x86_64.deb` 2,247. The Arch `.pkg.tar.zst` got 703 across both arches.+- v0.2.8 per asset: setup.exe 244, apk 163, msi 51, macos-arm64 45, AppImage 21+15, deb 16+3, ipa 10.+- **The in-app updater is barely used (V).** v0.2.7's `manifest.json`, which only the updater fetches (`src/crates/concat-host/src/updates.rs:417`), was downloaded 1,339 times against 126k binaries, about 1%. Updates are checked only when the user opens Settings › Version (`src/crates/concat/src/panes/settings.rs:509`). That fits the privacy policy, but nobody is told a new version exists, so each release has to win its audience back from scratch.+- Windows ARM was dropped after v0.2.6 (534 + 153 downloads that release; #326). The x64 setup still installs on ARM under emulation (`assets/windows/concat.iss:15`, `ArchitecturesAllowed=x64compatible`).++### Repo (V, `gh api repos/jub0t/Concat`)++9,638 stars, 1,161 forks, 79 watchers, 157 open issues+PRs (141 issues, 16 PRs), AGPL-3.0, created 2026-08-25. Homepage field: `concat.standalones.org`.++I sampled the last 100 stargazers (03:59–10:27Z today, about 100 stars in 6.5 h). 6 accounts were under 30 days old and 16 had no repos and no followers. That looks organic, not bought (V, sample only).++### Traffic, 14 days (V, `/traffic/*`)++- **Views:** 166,377, 91,011 unique. The peak was 26,800 on 10-16; daily views were above 19k on 10-15 to 10-18 and fell to 14,423 on 10-19.+- **Clones:** 11,098 (3,441 unique).+- **Referrers:** Google 41,694 (33,424 unique); github.com 21,441; Facebook (l./lm./www.) 11,202 combined; Brave 2,343; `concatenate.pages.dev` 2,228; Bing 1,909; t.co 1,219; youtube.com 1,180.+- **Paths:** `/jub0t/concat` and `/jub0t/Concat` are about 105k combined. `/releases` 15,520 and `/releases/tag/v0.2.7` 6,417. `docs/README.es.md` is the 10th most-viewed path (1,962 views), so Spanish speakers are a real segment.++## 2. Distribution and trust++### Where Concat is carried++| Channel | Status | Evidence |+|---|---|---|+| GitHub Releases | Yes, 16 assets / 5 OSes | V |+| Official site | `concat.standalones.org` (in-app `SITE`, `src/crates/concat/src/panes/start.rs:94`) and `concatenate.pages.dev` (README). Same HTML. **Both link only v0.2.7 assets** (47× "0.2.7", 0× "0.2.8"), plus `releases/latest/download/SHA256SUMS`, which is now 0.2.8's file. Anyone checking the hash of a site download gets a **mismatch**. Owner issue #458 ("website: update ... out-dated data") is open and unlabelled. | V (`curl`, 10:4xZ) |+| winget | **No** (`microsoft/winget-pkgs/manifests/j/jub0t` and `.../c/Concat` 404; code search empty). `concat.wxs:9` says the .msi is for "winget" deploys, but there is no manifest. | V |+| Microsoft Store | **No.** Removed in `be1b882` (2026-10-12, "no Microsoft Store"). Repo **variables** `MSSTORE_IDENTITY_NAME=jub0t.Concat`, `MSSTORE_PUBLISHER`, `MSSTORE_PUBLISHER_DISPLAY_NAME` are still set (`gh variable list`), so the name reservation probably still exists. | V |+| Google Play | No (`play.google.com/...id=app.concat.editor` 404) | V |+| F-Droid / IzzyOnDroid | No (both APIs 404) | V |+| Flathub | No (`flathub.org/api/v2/appstream/app.concat.editor` 404) | V |+| Flatpark | Yes, third party. Its page says "community package ... not affiliated", "repackages the official upstream Linux x86_64 .deb ... unmodified", still **0.2.7**, "127 new installs in 30 days". The README lists it with a ✅ Linux line. | V (WebFetch) |+| Homebrew | No (`formulae.brew.sh/api/cask/concat.json` 404) | V |+| AUR | Community: `concat`, `concat-bin` (3 votes), `concat-git`, maintained by **Felitendo**, still at **0.2.7-1** (2026-10-16) | V (AUR RPC) |+| App Store / TestFlight | No. The `.ipa` is ad-hoc signed and needs re-signing by the user (`.github/workflows/mobile.yml:30,414-430`) | V |++### Code signing per platform (V: `gh secret list` shows only `ANDROID_KEYSTORE` and `ANDROID_KEYSTORE_PASSWORD`)++| Platform | State | Evidence |+|---|---|---|+| Windows | **Unsigned.** `build-app.yml:1032-1037` prints "No signing certificate: the build is unsigned" because `WINDOWS_CERTIFICATE_PFX` is unset. SmartScreen warns, and Smart App Control **refuses** to run it (CODE_SIGNING.md:3-6). SignPath is test-signing only (maintainer). CODE_SIGNING.md never mentions SignPath. | V |+| macOS | **Ad-hoc, not notarised**: `build-app.yml:698-699` `codesign --force --deep -s -` (no `APPLE_*` secrets). The README tells users to run `xattr -dr com.apple.quarantine`. | V |+| iOS | Ad-hoc; the user signs it with their own Apple ID (AltStore/Sideloadly). 10,279 people downloaded the v0.2.7 IPA. The site offers it to casual users directly, and #338 ("I can't download it on ios") is unanswered. | V |+| Android | Signed with the repo key: `CN=Concat, O=Concat`, cert SHA-256 `923308579c0b6f4d6d9be39c504f5f05f248feeae41d5caece23f058a29d6e2e`, v2+v3, for v0.2.6/0.2.7/0.2.8. **v0.2.5 was signed with a different key** (`4e6bdfb2…b042`), so the 15,206 people who installed v0.2.5 cannot update in place: Android refuses a different signer, and they must uninstall and lose app data. No release note found warns about this (U whether any was ever posted). | V (`apksigner verify --print-certs` on all four APKs) |+| Linux | Unsigned packages. SHA256SUMS is unsigned (no GPG/minisign/cosign). | V |+| All | **No build provenance or attestations**: no `attest`/`sbom`/`cosign`/`id-token` anywhere in `.github/workflows/`. The updater checks bytes against a digest in `manifest.json` from the **same release** (`updates.rs:6-13`), so one compromised GitHub token or account can ship code to every updater. On the positive side, all third-party actions are SHA-pinned. | V |++### Android developer verification readiness++- Google's schedule: enforcement for **Brazil, Indonesia, Singapore and Thailand from 2026-09-30**, global in **2027**. On certified devices, normal installs of apps from unregistered developers are blocked; adb and a slowed "advanced flow" remain ([developer.android.com](https://developer.android.com/developer-verification/guides?hl=en); [The Hacker News](https://thehackernews.com/2026/06/google-sets-sept-30-deadline-for.html?hl=en); [Help Net Security](https://www.helpnetsecurity.com/?p=375237)). Whether the 30 September enforcement actually started on time is **U**: I found no post-deadline reporting.+- Concat has **no sign of registration**. The repo, docs and issues contain no mention of developer verification (V, grep).+- **It is feasible:** the package name is `app.concat.editor` and the key has been stable since 0.2.6 (V), which are the two things verification ties to an identity.+- **Exposure:** Indonesia is plausibly a real market. There is an Indonesian locale, Indonesian contributors (PR #383 Indonesian TTS, @dapaaa07) and an Indonesian README. Country-level download data does not exist, so the size of the impact is **U**.+- Other Android gaps (V, `aapt2 dump badging`):+ - The APK is **arm64-v8a only**, so 32-bit phones (#392, unanswered) and x86_64 Chromebooks are out. Drift ships arm64, armeabi-v7a and x86_64.+ - targetSdk 35. That is fine for sideloading, but Play would want 36 after Aug 2026 (U, based on Google's annual rule).+ - 0.2.8 adds `RECORD_AUDIO` (voiceover). Otherwise it only asks for `INTERNET`.+ - The release notes admit "no automated check has started this APK": CI's emulators are x86_64 (`mobile.yml:34-43,106`).++### Installer behaviour (Windows)++- `assets/windows/concat.iss`: per-user by default (`PrivilegesRequired=lowest`, can be raised from a dialog), 25 installer languages vendored, desktop icon opt-in, launches after install.+- `assets/windows/concat.wxs`: per-machine (`Scope="perMachine"`), Start menu only, no UI, and no desktop icon.+ - The two installers have different AppId/UpgradeCode values. Installing both gives **two copies and two "Concat" Start entries** (U, by construction).+ - The updater does pick the right installer per install (`updates.rs:280-290,491-504`, V).+- **"Concat (safe graphics)"** (`64e9dcd`, 2026-10-20 09:27Z, added to both installers):+ - It is **not in any release**. It landed an hour after v0.2.8 was published, and the tag is `79116bd`. Users hitting #389/#449 (garbled window) have nothing in hand today.+ - It is **sticky**. The commit says the entry "turns the switch off before the window opens and keeps it off." One click leaves a GPU video editor on the software rasteriser forever, including from the normal shortcut, until the user finds Settings › General › Draw on the GPU. Expect "Concat is slow" reports that are really this setting.+ - The shortcut is always created (not optional), so every install gets a second, confusing Start entry.+- Separately, `b18c2fe` (post-tag, #418) pins D3D12 to **FXC**. That partly reverses `be1b882`, which moved to DXC because FXC "takes minutes over the Gaussian loops of the blur and the glow" (#297). That trade-off needs proving before v0.2.9 (U; the code is outside this area).++### Impersonation and lookalikes (flag only, nothing done)++- **`github.com/Concat-video-editor`** is an Org created 2026-10-16 (the day of the traffic peak).+ - Its `.github` profile README is SEO filler: "Concat - Online Video Editor for Fast Browser Editing", "a practical Concat CapCut alternative".+ - It **hotlinks `concat.standalones.org/_astro/editor…webp`** and gives no link to `jub0t/Concat`.+ - This is name squatting or SEO spam riding Concat's name (V, `gh api orgs/Concat-video-editor`). `TRADEMARK.md` exists, and the owner can report it to GitHub.+- There are **two official domains** (`concat.standalones.org` canonical, `concatenate.pages.dev` in the README). Two names make phishing lookalikes easier to pass off; pick one.+- README "Paid Sponsors" promotes Proxyon (proxies, `JUB0T` code). That is a trust-optics risk for an app sold on privacy. This is opinion.++## 3. GitHub state++### Issues (V, GraphQL over all 141 open)++- **Labels:** feature request 70, bug 51, testing-requested 32, ui-ux 27, important 24, windows 22, priority 19, timeline 17, video 16, crash 15, effects 14, android 13, audio 13, captions 12, improvement 12, gpu 10, AI 9, export 8, ios 8, editor-api 7, distribution 6, linux 6, text 6, needs-info 5, performance 5, macos 4, plus smaller ones. 4 are unlabelled: #455, #457, #458 (owner) and #459 (outside, today).+- **Per platform (open):** Windows 22 (18 bug, **10 crash**), Android 13 (5 bug), iOS 8 (3 bug, 1 crash), macOS 4 (2 crash), Linux 6 (1 crash).+- **Authors:** 33 owner-authored, 108 by others.+- **No maintainer reply: 70 of 108 outside issues (65%).** The owner wrote 77 of all comments on open issues.+ - The oldest unanswered issue is #155 (2026-09-19, 31 days). The oldest with zero comments at all is #200 (09-25).+ - **Oldest unanswered crash: #294 "App crash at launch in IOS"** (10-09, 0 comments).+ - Also unanswered: #313 (AVX2 launch crash, 10-12), #324 (Linux Vulkan export device loss, priority), #343 (Ubuntu preview won't play), #371 (Intel Xe renders no UI), #423 (macOS legacy AMD launch failure), #437 (⌘Q loses edits), #449 (green screen, Windows), #338 (iOS can't download).+- **`testing-requested` (32): the process is broken.**+ - The labels went on today between 08:41 and 09:56Z, after v0.2.8 was published (e.g. #313 at 09:55:53Z, #372 at 08:42:29Z, via the timeline API).+ - **14 have zero owner comments** (#213, #239, #244, #253, #286, #313, #347, #350, #355, #356, #372, #413, #415, #416). A label change does not notify the reporter (U, but GitHub does not send notifications for label events), so these reporters have not been asked to test.+ - **#415, #416 and #418 have their fixes only after the tag** (`a1fd411`, `f2d0637`, `b18c2fe`/`a9b2df6`). No downloadable build contains them, and #418 is a Windows startup crash.+ - #86, #395, #413 and #434 have no commit that references them.+ - #223 and #297 were confirmed by their reporters (R) and are still open.+- **Discussions:** 29 in total. 8 of the 12 newest have 0 comments (D#398, D#387, D#384, D#376, D#366, D#298 …).++### PRs (V, `gh pr list`, reviews)++- **16 open.** Reviews posted today 10:04–10:09Z are substantive and correct. For example, #383 was blocked because the Piper archive's `MODEL_CARD` describes a Malayalam voice, so its `licence = "MIT"` has nothing behind it. #360 was told to drop the dead `vn.json`.+- **Approved but unmerged:** #446, #445, #441 (full CI green: 10 checks), and **#424 and #420 approved with only GitGuardian run, no CI**. CODE_SIGNING.md promises review before merge; review happened, testing did not.+- **Stale contributor work:**+ - **#157** (Mehdidjah, "geometric masks and motion tracking", +4,494, opened 09-19) has **zero reviews in 31 days**. Meanwhile the owner shipped his own shape masks in 0.2.8 (#239), which silently supersedes it.+ - #358 (template gallery, draft since 10-15). #425/#426 (MCP drafts, 10-18).+ - Dependabot #411 (21 cargo bumps) and #368 (6 action bumps) are unreviewed.+- **Closed PRs:** 117, of which 71 merged and 46 not. 43 of the unmerged are outside contributors'. #300, #187 and #275 were closed with **no comment or review at all**.+ - #300's work was ported as `b12e527` with authorship and sign-off kept (good), but the PR itself got no word.+ - The v0.2.8 release notes have **114 bullets and zero contributor credits** (no `@` in the body).+- **Git authorship:** 936 + 55 commits by Jareer. Co-author trailers: 349 Claude Fable, 255 Claude Opus 5.5, 14 Cursor.+- **Translations:** 133 of about 157 commits to `src/crates/concat/locales/` are by Jareer. They are effectively machine translations with little native review (U for quality).++### Trolling, spam, prompt injection (flag only)++- I scanned every open issue, comment and PR body for injection phrasing, hidden HTML, shorteners, file-sharing links and crypto terms. **Nothing hostile found.**+ - Hits were benign: users saying they used ChatGPT or Claude (#131, #318, #389), pasted HTML (#388), and dependabot bodies.+ - Attachments are `.log`/`.txt` files on `github.com/user-attachments` (#388, #389, #395, #449). Treat them as untrusted when opened.+- Low-value or vague issues: #307 "Feedback", #362 "features request", #393 "transition add", #408 "Request New Feature".+- **Owner's own #363** proposes to "reverse engineer and allow using assets from already installed CapCut". That is a copyright and ToS liability; ByteDance owns those packages. Flagged for the owner.+- External squatting: see `Concat-video-editor` above.++## 4. Product claims vs reality++README.md at HEAD; release notes from the v0.2.8 body.++| Claim | Verdict | Evidence |+|---|---|---|+| About: "for macOS, Windows, Linux and Android"; Highlights: "macOS, Windows, Linux, Android" | **Inconsistent.** Platforms lists "✅ iOS / iPadOS", and the site serves the IPA (10k downloads). | README.md:46,77,92 |+| "✅ Windows" | **Overstated.** Unsigned (Smart App Control blocks it outright), 10 open crash issues, and the #418 startup crash fix is in no release. | §2, §3 |+| "✅ Android · phones and tablets" | **Overstated.** arm64 only; the release notes say no automated check has launched the APK; the GL ES fallback is "only run on an emulator". | notes "Before you install"; `aapt2` |+| "✅ iOS / iPadOS · sideloaded" | **Overstated; should be 🧪.** Ad-hoc IPA, #294 launch crash unanswered, #338 unanswered. | §2, §3 |+| "✅ Linux ... Also a Flatpak on Flatpark" | True but third party, unaffiliated, and still 0.2.7 | Flatpark page |+| "4K scrubs smoothly" | **Unverified, and contradicted for many users.** No 4K or scrub benchmark or test in the repo (grep of `crates/*/tests`, no benches). Open #261 "Massive Performance Issues", #149, #412, #343. Android decoded on the CPU until 0.2.8 ("Video decodes on MediaCodec, which it never did"). | V |+| "100 effects, filters and transitions" | **True by count:** 23 effect + 57 filter + 22 transition = 102 visual packages (+25 audio) in `src/crates/concat-effects/packages/*/effect.toml`. **But in v0.2.8, 17 of 22 transitions declare knobs no user can set** (V: 15 have 1 param, 2 have 2). That was fixed only post-tag in `ce64dae`. The "just a fade" test `4451548` is also post-tag (#448). | V |+| "25 languages" | **True by file count:** 25 locales × 882 keys, none missing. **Overstated in practice:** de 87, fr 93, it 67, es 63 strings identical to English (some legitimately); hard-coded English (#352, PRs #424/#360); effect names English-only by design (#364); no UI mirroring for ar/he/fa/ur found (bidi exists only in caption text, `concat-text/src/lib.rs:983-1010`). | V (script over `locales/*.json`) |+| "100% local. Nothing uploads." / "no telemetry" | **Consistent with code checked.** Update check is manual; models are fetched on request; start-page cards are links. | `updates.rs`, `settings.rs:509`, `start.rs:94-110` |+| "Export H.264, HEVC, AV1. Up to 4K 60, 10-bit" | **Partly verified.** Codec names and HDR 10-bit are in `concat-export/src/lib.rs:418-486`. Per-platform encoder availability is U. #286 (Android missing codecs) is testing-requested with no reply. | V/U |+| "Auto-captions. Local Whisper ... styled" | Works, with accuracy caveats: #438 captions land seconds early; #325 asks for a larger model. | V (issues) |+| "Same app, same project files" | U | |+| "Download step 1: The website hands you the right build" | **False today.** The site hands out 0.2.7 and a mismatched checksum link. | §2 |+| Version badge "0.2.8" | True for GitHub; false for the site | |+| Notes: "Q and E cut the selected clip at the playhead" | **False in 0.2.8.** It was superseded by `dce4ad0`: Q = delete left, W = delete right, **E unbound** (`keymap.rs:116-120`, tests at `:613-615`). docs/KEYBOARD_SHORTCUTS.md:26-27 is correct. The notes contradict themselves. | V |+| Notes: "Intel Macs open on macOS 12 again (#372…)" | **Unverified.** Reporter never contacted, labelled testing-requested today. | §3 |+| Notes: "Delete left/right … (#400)", "Stabilize" (#253), masks (#239) | Code is in the tag (V); no reporter confirmation (U) | |+| Notes download table | Omits the shipped Arch `.pkg.tar.zst` | V |+| Notes as a whole | 114 bullets, many internal ("the lock covers one far byte", "uniforms come in whole 16 bytes", "Rational::parse refuses a value past i64"). Unreadable for the CapCut-refugee audience. | V |+| CODE_SIGNING.md: "Every change from anyone else ... is reviewed by a reviewer before it is merged" | True for review. But #424/#420 were approved without CI, and "Every member uses multi-factor authentication" is U. | V/U |+| SECURITY.md: "Report it privately through GitHub's advisory form" | **Broken.** `gh api repos/jub0t/Concat/private-vulnerability-reporting` → `{"enabled":false}`. | V |++## 5. Competitive position++| | Concat | CapCut | Drift (CutWire-Studios) | OpenCut | Kdenlive / Shotcut |+|---|---|---|---|---|---|+| Licence / stack | AGPL-3.0, Rust + Slint + wgpu (V) | Proprietary (ByteDance) | GPL-3.0, C++/Qt 6 + FFmpeg (V) | MIT, TypeScript web, Rust rewrite in progress (R) | GPL-3.0, C++ (V) |+| Stars | 9,638 (V) | n/a | 2,039 (V, created 2026-07-07) | 95,433 (V) | 5,904 / 15,456 (V) |+| GitHub downloads | ≈204.9k all-time; 126k for v0.2.7 (V) | n/a | 94,777 all-time; v0.8.0 9,702 (V) | Releases are FFmpeg builds only (V) | Kdenlive 11,500,714 official downloads in 2025 (V, [kdenlive.org](https://kdenlive.org/news/2026/state-2026/)) |+| Platforms | Win x64, macOS, Linux, Android arm64, iOS ad-hoc (V) | Win, macOS, Android, iOS, web | Linux, Windows, macOS (Apple silicon), Android arm64/armv7/x86_64 (V, assets) | Web (R) | Desktop only; Kdenlive mentions no mobile app (V) |+| Stores / package managers | **None first-party**; AUR and Flatpark community, both 0.2.7 (V) | All major stores | **Flathub (16,880 installs, V), Microsoft Store (listing 301 → V exists, U content), Homebrew tap (V), Play closed testing (Drift README, U)**, portable zip (V) | n/a | Kdenlive: Flathub etc.; Microsoft Store "soon" (V) |+| Signing | Unsigned Windows, ad-hoc macOS (V) | Signed | Ad-hoc macOS by its own README (V); Windows via Store (U) | n/a | Signed Windows (U) |+| Price / catches | Free, no account, no watermark (V) | Pro $179.99/yr, up from ~$77–90 ([Newsweek](https://newsweek.com/app-used-millions-nearly-doubles-subscription-price-overnight-11535999), V across sources); content-licence terms (R) | Free, no account, no watermark (Drift README) | Free | Free; Kdenlive €9,344.80 donations, 2 developers (V) |+| AI | Local Whisper captions, TTS + voice cloning, cutout, enhance (V code) | Cloud plus on-device | Auto captions, upscale, face retouch (Drift README, U) | U | Shotcut Whisper; TTS needs Docker (R) |+| Automation | MCP + JSON-RPC (V) | No | U | U | MLT XML |++Read-across:+- Drift is the direct threat on distribution, not features. It is smaller and younger, yet installable from three stores and is first in line on Android.+- The report's claim that there is "no open-source rival with a mobile app" (`reports/What Concat users want most.md`) is **false**: Drift ships Android APKs (V).++## 6. Scores++| Aspect | Score | Justification |+|---|:-:|---|+| User likeability | **5** | Real CapCut-parity gains in 0.2.8 (delete L/R, masks, stabilize, SRT, voiceover) and huge demand. But 65% of outside issues are unanswered, 10 Windows crash issues are open, iOS crash reports go unanswered, and the Android APK has never been launched by CI. |+| Documentation | **5** | Effect and locale counts are honest and the release-note caveats are candid. But SECURITY.md points to a disabled channel, the README platform claims contradict each other, "4K scrubs smoothly" has nothing behind it, the website is a release behind with a wrong checksum link, and the notes are 114 jargon bullets including a false keybind. |+| CI and release (distribution angle) | **4** | 16 artifacts across 5 OSes with SHA256SUMS and SHA-pinned actions. But the website was not updated with the release, testing-requested fixes are missing from the build, there is no notarisation, no signed checksums and no provenance, and two PRs were approved with no CI. |+| Security (supply chain / licences / trust) | **4** | Dependabot, push protection, model licences declared, and a sharp licence review on #383. Against that: unsigned Windows, ad-hoc macOS, an updater trusting a same-release digest, no attestations, a disabled vulnerability-reporting channel, and an Android key rotation that stranded 0.2.5 users. |+| **Distribution and trust (new)** | **3** | GitHub plus a stale site only. Not on winget, Store, Play, F-Droid, Izzy, Flathub or Homebrew; community AUR and Flatpark lag a release. Unsigned or ad-hoc on three OSes, no Android developer verification while enforcement has started in four countries, and an SEO squatter using the name. A younger rival is in three stores. |++## Highest-value fixes (ordered)++1. **Today:** point the site's download links at v0.2.8, or better, make them derive from `releases/latest`, and fix the SHA256SUMS mismatch.+2. **Close the `testing-requested` loop.** Post one comment per issue saying "fixed in 0.2.8, please test on <asset>". Un-label #415, #416 and #418 until a build contains their fixes.+3. **Turn on private vulnerability reporting** (Settings › Security). It is a single toggle.+4. **Register as an Android developer** for `app.concat.editor` with the current key. Submit to IzzyOnDroid, which takes developer-built APKs.+5. **Get onto the stores.** Microsoft Store registration is free for individuals and Microsoft signs Store apps (R, [Windows Developer Blog](https://blogs.windows.com/windowsdeveloper/2025/09/10/free-developer-registration-for-individual-developers-on-microsoft-store/)); the `jub0t.Concat` identity is still configured. Then a winget manifest for the .msi and a Flathub submission (Drift proves both work for this category).+6. **Fix the README platform marks.** Mark iOS 🧪, add Android "arm64 only, untested on real devices in CI", and drop "4K scrubs smoothly" until a benchmark exists.+7. **Make "safe graphics" one-shot**, or show a banner while software drawing is on. Ship it in v0.2.9 with a note.+8. Publish signed checksums (minisign or cosign) and GitHub artifact attestations. Have the updater verify a signature, not a same-origin digest.+9. Credit contributors in the release notes, and say a word when a PR is closed or superseded (#157, #300).+10. Report the `Concat-video-editor` org to GitHub, and choose one canonical domain.diff --git a/audit/2026-10-20/project-effects-ml-phones-ci.md b/audit/2026-10-20/project-effects-ml-phones-ci.mdnew file mode 100644index 00000000..0a773f67--- /dev/null+++ b/audit/2026-10-20/project-effects-ml-phones-ci.md@@ -0,0 +1,156 @@+## project, effects, speech, vision, Android, iOS, CI and release: HEAD ce64dae (2026-10-20)++Baseline: audit/2026-10-15 (HEAD ab21457). Read-only. No cargo, no builds. Evidence = source at ce64dae (`git show ce64dae:<path>` where the working tree differs), CI/release via `gh` (read-only), and the **shipped v0.2.8 APK**, downloaded and inspected by script (ELF headers, binary manifest). Things not run on a device are marked *unverified*.++Note for the coordinator: during this audit `main` moved to baaeb2a (locale fix) and the working tree holds **uncommitted** edits to `studio.rs`, 25 locales and `media-pane.slint` (someone is fixing N4/N5 below live). Findings cite ce64dae, not the working tree.++---++### 1. Re-verification of the 10-15 findings++| # | 10-15 finding | Status at ce64dae | Evidence |+|---|---|---|---|+| 1 | Release notes' previous tag | FIXED (holds) | 0.2.8 notes span 0.2.7..0.2.8 |+| 2 | iPhone Import dead (#332) | FIXED in code, **unverified on device** | 8dfa7bb; `concat/src/ios.rs` (660 lines) installs PHPicker + UIDocumentPicker, `adopt` :516. Simulator smoke never opens a picker. #332 still open |+| 3 | Phones not built on push | PARTLY | push trigger added `mobile.yml:77`; but `cancel-in-progress` for push `mobile.yml:92` → of the last 60 main runs since 16 Oct: **50 cancelled, 7 failed, 2 succeeded** (both on the 0.2.8 tag commits 9f0cf73, 79116bd). Main is effectively still not phone-built |+| 4 | Windows heavy GPU suites skip on WARP as pass | UNCHANGED | `concat-render/src/gpu/tests.rs:41-47` (`heavy_gpu`). Effects still never run through D3D12 in CI |+| 12 | Phone that fails to start vanishes; iOS plist keys | PARTLY | plist keys FIXED `src/scripts/ios-app.sh:99-100`; Android still `return`s silently `concat-android/src/lib.rs:519-521` |+| 13 | `locales.py --check` passes with missing keys | UNCHANGED | `scripts/locales.py:193-194` prints `missing:` without `failed = 1` |+| 14 | CI internals in release notes | PARTLY | caveats section new (`.github/release-notes.sh:84-95`), but 0.2.8 notes still lead with "No effect cards are drawn on a software adapter, which froze CI's API 30 emulator" and "the lock covers one far byte" |+| 15 | DirectML asked of CPU ORT | FIXED | e461750; `concat-vision/src/runtime.rs:183-190`, `concat-speech/src/lib.rs:28-36` |+| 16 | Intel Mac ORT minimum | PARTLY | ORT now `load-dynamic`, lazily opened `build-app.yml:313-330, 640-646`, `runtime.rs:64-85`. But see N2: FFmpeg dylibs in the same bundle need 14.0. #372 open, no reporter confirmation |+| 17 | `--smoke` stops before the window | UNCHANGED | `concat/src/lib.rs:2362-2385` exits before `app.run()` (:2419); `smoke-windows.ps1:17-18` still claims "through the first frame" |+| 18 | ISA guard | PARTLY | SDE now also runs `--whisper-cpu` under `-snb` `smoke-windows.ps1:190`; Ivy Bridge still `--version` only (:182). Allowlist widened for MSVC STL AVX2 `isa-allow.txt` (79116bd) - explanation plausible (dispatched `vector_algorithms`), counts hard-coded with margin; `ggml-cpu-ivybridge.dll` row added without a measured count |+| 19 | Bundle workflow tag-only | UNCHANGED | `build-app.yml` `on: workflow_call` only |+| 20/H2 | Windows unsigned; nothing stops it | UNCHANGED (now disclosed) | v0.2.8 log 03:23:05 "No signing certificate: the build is unsigned."; `build-app.yml:1036-1037` exits 0. Notes say so (`release-notes.sh:85-86`) |+| 21 | Dependabot cargo group red | UNCHANGED | #411 (21 updates) open; its phone run 39180952703 failed |+| 28 | Concurrency cancels main | PARTLY | `ci.yml:32` no longer cancels in progress, but GitHub keeps one pending run per group, so queued main runs are still cancelled: **22 of last 60** CI runs on main cancelled |+| 29 | `ubuntu-latest` | FIXED | 94d4777; no `ubuntu-latest` in `.github/workflows` |+| §7 | smoke exit/trap/exec bit | FIXED (holds) | |+| – | Phone downloads without digest | UNCHANGED | `src/scripts/ffmpeg-mobile.sh:39-40`, `sherpa-mobile.sh:42, 60` (prebuilt native code from k2-fsa, unverified, into the shipped APK/IPA) |+| – | Cards skip with no adapter | WORSE in coverage | d9f7e08 now also skips on *software* adapters (`concat-host/src/cards.rs`), so Android emulators never compile the effect shaders at launch |+| – | AddClip takes whole media | UNCHANGED | `concat-project/src/commands/clips.rs:24-46` |+| – | Speed curve grows over neighbour | UNCHANGED | `commands/properties.rs:309-330`: duration recomputed, no ripple, no overlap check |+| – | All project tests in one file | WORSE | `concat-project/src/lib.rs` 4 533 → **5 157** lines, 136 of the crate's 150 tests |+| H1 | main red, Linux tests skipped | RECURS | main red again since 08:39Z today: 6599b5a build timed out at the 25-min cap (`ci.yml:160`), then be60358/a1fd411/f2d0637 failed `locales.py --check` (`ci.yml:141`) after f121747 removed `sounds.*` UI without its keys - clippy (:151) and tests (:168) skipped. 17-18 Oct: 22 consecutive failures (fmt, locales, tests). **28 failed + 22 cancelled of 60** main CI runs since 16 Oct; greens cluster on release commits only |+| H3 | Android check never ran app code | PARTLY FIXED | 78e4be1/766fd4e: emulator now installs an **x86_64 build of the same source** and it really starts (v0.2.8 API 30 & 35 logs: "Concat 0.2.8 starting", ResumedActivity = app). The shipped arm64 APK is still never started; the x86_64 test build links sherpa's ORT, not the shipped one, so the real-device ORT start path is still untested. Notes say so (`release-notes.sh:88-89`) |+| H4 | Intel Mac 12.0-13.2 crash | PARTLY | see 16 and N2 |+| H5 | Activity recreation kills app | FIXED (verified in shipped manifest) | 887a034, `concat-android/Cargo.toml:74`; shipped APK `configChanges=0x50003ff4` = all 13 flags. Still missing `mcc|mnc|touchscreen|colorMode|grammaticalGender` (rare). Second-`android_main` guard moot: the loop end calls `leave()` → `process::exit` (`concat/src/lib.rs:2419-2434`) |+| M1 | Import crash after process kill / provider error | PARTLY | SecurityException FIXED (724a256, `ConcatFiles.java:355-364` catches `RuntimeException`). Dex-loaded Fragment restore after process death UNCHANGED (`lib.rs:154-195`, `ConcatFiles.java:221-225`: system restores `app.concat.editor.ConcatFiles` with the activity's class loader → ClassNotFoundException). *Reasoned, unverified* |+| M2 | Android projects deleted on uninstall | UNCHANGED | `lib.rs:37-38` HOME = external_data_path; imports copied to `getExternalFilesDir/Imported` (`ConcatFiles.java:282`), no de-dup (`unique()` :325-340 makes "name (2)") |+| M3 | 0.2.7 notes false claim | FIXED | v0.2.7 body now starts "Corrected on 16 October" |+| M4 | Shipped .ipa never launched; prebuilt minimums | PARTLY | iOS min-OS check added `mobile.yml:420-425` (passes: "2 Mach-O files... every file runs on 15.0"); still a separate sim build on newest runtime only; #294 open |+| M5 | Nobody knows if main builds for phones | PARTLY | see 3 |+| M6 | Vision: no CPU retry on accelerator run-time failure | UNCHANGED | `runtime.rs:94-110, 140-143`; NNAPI/CoreML never in CI |+| M7 | No offline shader portability check | UNCHANGED | `concat-effects/Cargo.toml` naga `wgsl-in` only |+| L1 | `pow` negative base | UNCHANGED | `packages/concat.light-leak-flash/effect.wgsl:17` |+| L2 | No `permissions:` | UNCHANGED | only `models.yml:45`, `release.yml:120`, `nix.yml:40` have one |+| L3 | Tag re-pointed after a failed run | REPEATED | v0.2.8 run 39180271851 on 9f0cf73 failed (Windows ISA scan), tag moved to 79116bd, run 39188732930 passed |+| L4 | Consolidation orphans renders | UNCHANGED (no commits) | `concat/src/studio/consolidate.rs` |+| L5 | Font mmap SIGBUS | UNCHANGED | `concat-text/src/lib.rs:654` |++Scorecard of 10-15 items: fixed 6 (1, 15, 29, H5, M3, §7), partly 14, unchanged 13, worse/recurring 4 (H1, project test file, card coverage, L3).++---++### 2. CI and release status (real, 20 Oct ~10:35Z)++- **v0.2.8**: tag 79116bd, built 19 Oct, draft published 20 Oct 08:18Z. All 16 jobs green incl. CI-in-release, 5 desktop bundles, Android arm64+x86_64, two Android smokes, iOS arm64 + simulator smoke. Downloads so far: setup.exe 244, APK 163, macOS arm64 45.+ - Windows: **unsigned** (log). macOS: ad-hoc, not notarised (annotation). Android: signed with the real key (V3 signer SHA-256 9233...6e2e, same as 0.2.7).+ - **Release does not fail when unsigned** (`build-app.yml:1036-1037`); it only adds a caveat (`release.yml:172`, `release-notes.sh:84`). Same for macOS min-OS violations (warnings only, N2).+ - Android API 35 smoke passed with a warning "window in focus is ... Application Not Responding: nexuslauncher" (job 117867497573): a system ANR dialog over the app is accepted.+- **main**: red since 08:39Z today (see H1 row). CI on main since 16 Oct: 8 success / 28 failure / 22 cancelled. Phones on main: 2 success / 7 failure (all 4 inspected: API 30 emulator froze, 30-min timeout, before d9f7e08) / 50 cancelled.+- **Phone workflows on PRs**: yes, `pull_request:` (`mobile.yml:76`); owner-branch PRs complete (39268016072 etc.); first-time contributors' PRs sit at `action_required`.+- **Does CI prove each platform starts?**+ - Windows x86_64: imports resolve, `--version`, `--smoke` (stops before the event loop), SDE Ivy Bridge `--version`, SDE Sandy Bridge `--whisper-cpu`. **No window, no first frame, no real GPU, no D3D12 effect compile.** #339/#365-class failures invisible. Windows ARM not shipped (#326).+ - Android: x86_64 emulator build really starts on API 30/35 for 60 s. **Shipped arm64 never started; no real device; no 16 KB-page device (N1).**+ - iOS: simulator build really starts (newest runtime only). **Shipped .ipa never launched.**+ - macOS: `--smoke` on runners (15.x); nothing on 12/13/14 despite declaring 12.0 (N2).+- **Model mirror never exists**: `models.yml` has **never run** (no runs) and `gh api .../releases/tags/models-v1` → 404. Every model download's first source (`concat-host/src/models.rs:29-42`) is a dead URL; the app always falls through to third parties (GitHub releases of PeterL1n, danielgatis, k2-fsa; Hugging Face at pinned commits). Integrity holds (digest required, `models.rs:311-321`, test :495), availability does not. `HF_REPO = ""` (:72) so China users get only hf-mirror of HF upstreams.++---++### 3. New findings++#### HIGH++**N1. The shipped Android app library is 4 KB-aligned: it cannot load on a 16 KB-page phone.** *Verified on the v0.2.8 APK.* ELF PT_LOAD alignment: `libconcat_android.so` **4096**; `libonnxruntime.so`, `libsherpa-onnx-c-api.so`, `libc++_shared.so` 16384. Only FFmpeg gets `-Wl,-z,max-page-size=16384` (`src/scripts/ffmpeg-mobile.sh:92`); the Rust cdylib gets nothing (no RUSTFLAGS/link-arg in `mobile.yml` or `mobile-env.sh`). Scenario: a phone running a 16 KB kernel (Pixel 8/9/10 "16 KB" mode, devices launching on Android 16+) → `dlopen` refuses "program alignment (4096) cannot be smaller than system page size (16384)" → app closes at launch with nothing shown (lib.rs never runs); on Android 16 compat-mode devices, a warning dialog at best. Google Play has required 16 KB alignment since Nov 2025, so this is also a hard blocker for any Play listing. Fix: `-C link-arg=-Wl,-z,max-page-size=16384` for `aarch64-linux-android`, plus a CI check over every `.so` in the APK (same shape as `check-min-os.sh`).++**N2. macOS bundles declare 12.0 and ship Homebrew dylibs built for 14.0/15.0; the check only warns, and 0.2.8's notes claim Intel Macs open on 12.** *Verified in CI annotations; launch effect unverified (no macOS 12-14 runner).* v0.2.8 run 39188732930: arm64 bundle - libavfilter, libavdevice, libx264, libdav1d, libSvtAv1Enc, libvmaf, libmp3lame, libmpg123, libssl, libcrypto "needs 15.0, newer than the declared 12.0"; Intel bundle - same set "needs 14.0" (libssl/libcrypto 15.0). `check-min-os.sh:9-12` itself says dyld "loads it, and the process dies on the first symbol the older system lacks"; `build-app.yml:640-654` runs it without `--enforce` ("A warning for now"). `build-app.yml:96-97` says "Apple silicon Macs all run newer than this" - false (M1 Macs run 11-15). Release note "🐛 Intel Macs open on macOS 12 again (#372, #333, #342)" is unproven: the ORT crash is removed, the next missing symbol may be FFmpeg's. Fix: build FFmpeg for the floor (as phones already do) or raise `MACOSX_DEPLOYMENT_TARGET` to 14/15 honestly; then `--enforce`.++**N3. Auto-captions start seconds early (#438) and nothing in the pipeline addresses it.** *Verified by reading; reporter's numbers untrusted but consistent with the code.* `concat-speech/src/transcribe.rs:481-497`: greedy whisper with no VAD, no token timestamps, no split-on-word, no max segment length; `:544-545` takes segment timestamps raw. Whisper's segment start is the previous segment's end, so every pause is swallowed into the next caption (reported 1.1-7.1 s early). Captions are the most-used AI feature in this class of editor. Fix: whisper.cpp's Silero VAD (or an energy VAD on the 16 kHz buffer already in hand) to trim leading silence per segment, `set_token_timestamps(true)` and re-anchor each caption on its first token; a regression test with a synthetic tone-gap file.++**N4. CI is not a signal on main: 50 of 60 phone runs cancelled, 50 of 60 CI runs red or cancelled.** *Verified via gh.* Causes: `mobile.yml:92` cancels in-progress push runs (a 50-60 min build vs pushes every 5-20 min); `ci.yml:31` group keeps one pending, cancelling the rest; cheap gates first (`ci.yml:136` fmt, `:141` locales) abort the job before clippy (:151) and tests (:168) - the H1 shape again today (f121747 dropped `sounds.*` without regenerating locales, 4 consecutive reds); the 25-min build cap (`ci.yml:160`) fails a cold cache (6599b5a, plus a cache-save tar error). Releases pass only because `release.yml` re-runs CI on the tag. Net: nobody can say which main commit broke a phone. Fix: run fmt/locales/models in their own job; `if: !cancelled()` on clippy/tests; don't cancel push runs on main (or run phones on a schedule against the latest main and let it finish); raise the build cap or prime the cache.++#### MEDIUM++**N5. ce64dae's transition knob: one undo step and a full model rebuild per slider tick.** *Verified at ce64dae.* `set_transition_param` uses `self.apply` (`git show ce64dae:src/crates/concat/src/studio.rs:4306`), not the gesture-folded path the effect knobs use (`studio.rs:5593-5611`); `Param.changed` fires on every move (`ui/inspector/controls.slint:421`). Dragging Wipe softness 0→1 = dozens of undo steps and an info log line each. Publishing creates a fresh `VecModel` every time (`:9044`), so the slider row is rebuilt under the finger mid-drag (worst on a phone). An uncommitted working-tree fix switches to `apply_within("transition-knob:{clip}:{key}")`, but never calls `end_gesture`, so two separate drags of the same knob with nothing between fold into one step (`concat-project/src/editor.rs:277-283`). No test covers either.++**N6. Transition params bypass the NaN guard; a non-finite value makes the project unopenable after save.** *Verified by reading.* `Command::has_non_finite` checks only `tr.duration` (`concat-project/src/commands/mod.rs:1092-1096`), not `tr.params` (filters' params are checked, :989-990). serde_json writes NaN as `null`; `from_document` (`doc.rs:170-173`) fails the whole parse on `null` for an `f64` and returns None → "nothing to open". The UI path refuses non-finite (`studio.rs` `is_finite` check), so the reachable callers are Rust API users/tests; latent but one line to fix. Also no render-time clamp: `concat-effects/src/shader.rs:1673-1685` writes any value from the file into the uniform, so a hand-edited or API-written `direction: 1e9` reaches the shader unvetted (install vet only tries min/default/max).++**N7. Project files written by ce64dae+ lose their transition knobs silently in 0.2.8.** *Verified by reading.* `Transition` has no `deny_unknown_fields` and `extra` passthrough is top-level only (`doc.rs:358-362`), so 0.2.8 (desktop or a phone that has not updated) drops `params` on read and the next autosave writes the project without them; `DOCUMENT_VERSION` stays 1 (`doc.rs:41`). Backward read is fine (`#[serde(default, skip_serializing_if)]`, `model.rs:1311-1315`). Acceptable for an optional knob, but it is the first field whose loss changes the picture; there is no per-object passthrough or minor-version marker to warn.++**N8. Model mirror does not exist and has never been filled.** See §2. `models/manifest.toml:8-14` promises "stays reachable... under one roof"; reality: every download depends on five third-party hosts, two of them personal GitHub releases (`rvm_mobilenetv3_fp32.onnx`, rembg `v0.0.0`). One deleted release = a feature gone for every user, with a 404 retry first.++**N9. Phone native dependencies are fetched without digests into signed release artifacts.** `ffmpeg-mobile.sh:39-40` (ffmpeg.org tarball), `sherpa-mobile.sh:42, 60` (prebuilt `.so`/xcframework from GitHub). Desktop downloads are digest-pinned (`scripts/ci/deps.lock`, `fetch.sh`); phones are not, and the APK is then signed with the real key. A compromised upstream release ships under Concat's signature to 16 000+ Android users.++**N10. Captions ignore the speed curve and the clip moving during the run.** `concat/src/panes/captions.rs:496` (window = duration × mean speed), `:523` (`speed: clip.speed`), `:211-212` (place at `start + t / speed`). On a ramped clip the captions drift by the integral difference; a clip moved or trimmed during a multi-minute transcription gets captions at its old position. *Reasoned.*++**N11. API 35 smoke accepts a system ANR over the app; nothing checks a picture.** `smoke-android.sh:322-324` turns "focus is not the app's" into a warning; the screenshot (`:102`) is collected, never inspected. v0.2.8's pass had the launcher's ANR dialog on top. A black or frozen first frame passes.++#### LOW++- **L-a** `AddClipsInSequence` non-magnetic can reverse order: each item takes `nearest_free_start` from the previous end (`commands/clips.rs:68-118`, `placement.rs:42-66`), whose candidates include `clip.start - duration` of the clip just placed. Lane with X at [26,60], drop A(5 s), B(5 s) at 20: A→20, B→15. 38056f1's "end to end on one lane" holds only when the gap is free. *Reasoned.*+- **L-b** SRT import is all-or-nothing: one malformed cue (`srt.rs` `parse`, `fraction.len() > 3`, missing blank line) rejects the file; `<i>`, `{\an8}` tags become literal caption text.+- **L-c** Voice cloning ("Clone from a sample…", `locales/en.json:550`) has no consent prompt and no watermark (`THIRD_PARTY_NOTICES.md:158-169` acknowledges both). Product/abuse risk, not a licence breach.+- **L-d** `orientation = "portrait"` (`concat-android/Cargo.toml:80`) and targetSdk 35 (:54): Android 16 ignores orientation locks on large screens for targetSdk 36; the phone shell will be forced into landscape on tablets/foldables once the target is bumped.+- **L-e** `ConcatFiles.done` uses `getExternalFilesDir(null)` unchecked (`ConcatFiles.java:282`); null on unmounted external storage → `mkdirs` on a relative path fails → the pick returns nothing, silently.+- **L-f** Card thread starved a 4-core emulator to a system freeze (d9f7e08 message); the fix hides it on software adapters only. Low-end real phones run the same 100+ pipeline compiles at launch. *Unverified on device.*++---++### 4. Effect package format, against the CapCut report++Verified at HEAD: one WGSL file + `effect.toml` + `fixtures.toml` + optional `.cube`; ≤8 passes (`manifest.rs:453`); unknown manifest fields denied; bindings fixed by the host; no image/sequence/video inputs, no mask input, no frame history, no audio/beat; unbounded loops refused at load (`shader.rs:1067-1113`), GPU vet at install. ce64dae delivered half of the report's Step 0: transitions now carry `params` (`model.rs:1311-1315`, threaded through `concat-export/src/lib.rs:384-387, 661, 810`), but not `keys` (knobs can't be keyframed), and no `clip_duration`/`frame_index`/`seed` in either uniform (`shader.rs` `TRANSITION_HEAD` :181: size, time, progress only). The sandbox is sound for security; it is the reason packages can't make the motion/flash/overlay effects users want. M7 (no offline HLSL/GLSL cross-compile) becomes mandatory before any binding is added.++---++### 5. Numbers (all `.rs` lines; tests = `#[test]`)++| crate | lines (10-15) | tests | unsafe | unwrap+expect | notes |+|---|---|---|---|---|---|+| concat-project | 12 899 (10 685) | 150 (121) | 0 | 446 | `lib.rs` 5 157 lines, 136 tests |+| concat-effects | 7 355 (7 256) | 72 | 0 | 139 | 127 packages, 1 966 WGSL lines |+| concat-speech | 3 504 (3 411) | 22 | 5 | 23 | `tts.rs` 1 504 |+| concat-vision | 5 498 (2 760) | 59 (31) | 0 | 32 | +stabilize |+| concat-android | 595 + 373 Java | **0** | 9 | 9 | |+| concat/src/ios.rs | 660 (new) | 0 | many `msg_send!` | – | |++---++### 6. Scores (1-10; previous from the 10-15 crate rows / overall table)++| Aspect | 10-15 | Now | Why |+|---|---|---|---|+| Code quality | ~7.3 | 7 | Careful, commented, tested commands; but ce64dae shipped the knob without the gesture path or NaN guard every sibling has (N5, N6) |+| Design architecture | ~8 | 7 ↓ | Full-snapshot undo and doc migration still good; transition params bolted on without keys or passthrough (N7); captions placement ignores the curve (N10); package sandbox sealed off from motion/masks |+| Design philosophy | ~7 (overall 5) | 6 | Honest "Before you install" caveats are a real step up; but gates that only warn (min-OS, signing, foreign ANR), merging on red, tag re-pointed again |+| Maintainability | ~6.5 | 6 | project test file +624 lines; studio.rs keeps absorbing features (knobs, iOS); Android 0 tests |+| User likeability | 6 | 6 | iOS Import, Android config changes, transition knobs, Intel ORT lazy: real gains. Captions early (N3), unsigned Windows, 16 KB phones can't open (N1) |+| Performance | ~5.8 | 6 ↑ | whisper leaves two threads, AVX2 back on capable CPUs, cards pre-shipped; nothing measured on a phone |+| Scalability | ~5.6 | 5 | per-tick republish of whole models; caption/undo snapshots of whole project; mirror absent |+| Security | ~7 | 6 ↓ | Model digests fail closed (good); phone deps unverified into signed APK (N9), Windows unsigned again, no `permissions:` in three workflows, no render-time clamp (N6) |+| Testing | ~5.7 | 5 | Android emulator smoke is now real (+), iOS min-OS check (+); but shipped arm64/ipa never started, WARP skips heavy GPU suites, main mostly red/cancelled, new knob path untested, Android crate 0 tests |+| Documentation | 6 | 6 | Notes corrected and caveated; still false: `smoke-windows.ps1:17-18`, `build-app.yml:96-97`, `models/manifest.toml:8-14`, "Intel Macs open on macOS 12 again" |+| **CI and release** | 4 | **5 ↑** | Real Android launch (x86_64), phones in release, min-OS scan, honest caveats. Held back: 50/60 phone and 50/60 CI main runs not green, warnings don't gate, unsigned release, tag moved, mirror never run, no 16 KB check |++Phone launch checks: Android 2 → 5 (real launch, wrong ABI); iOS 7 → 7.++### 7. Top fixes, in order+1. 16 KB link flag for the Android cdylib + an alignment check in `mobile.yml` (N1). An hour.+2. Decide the macOS floor honestly: build FFmpeg for 12, or declare 14/15; then `check-min-os.sh --enforce` (N2).+3. Captions: VAD/leading-silence trim + token timestamps, with a gap test (N3, #438).+4. Split lint gates into their own job, stop cancelling main push runs, raise the cold-cache cap (N4).+5. Commit the knob gesture fix with an `end_gesture` timer like the effect knobs; add `tr.params` to `has_non_finite`; clamp at `lay_params` (N5, N6).+6. Run `models.yml` once (N8); pin phone downloads in `deps.lock` (N9).+7. Fail a tag build without a Windows certificate, or keep the caveat and stop calling the release a pass.diff --git a/audit/2026-10-20/window-subreviews.md b/audit/2026-10-20/window-subreviews.mdnew file mode 100644index 00000000..c25ffd1e--- /dev/null+++ b/audit/2026-10-20/window-subreviews.md@@ -0,0 +1,121 @@+# Window sub-reviews, 20 October 2026++Four read-only reviews the window reviewer ran alongside+[`window.md`](window.md), which reported partial without them. Code at+`baaeb2a`. "Verified" means traced in the source, including Slint 1.18.1+and winit 0.30.13 in the cargo registry; nothing here was run on a device.+Where a finding was fixed later today, the fixing commit is named.++## 1. Today's commits++| Commit | Finding | Severity | Status |+|---|---|---|---|+| ce64dae transition knobs | Knob rows were a fresh `ModelRc` on every publish (`studio.rs:9044-9045`), and Slint drops every repeated row when the model handle changes (`i-slint-core` `model/repeater.rs:663-672`), so a scrub stopped after one step | High | Fixed in d4ced41 (kept model + `sync`) |+| ce64dae | One undo step per value change (`self.apply`), up to 100 steps a drag against `UNDO_DEPTH = 200` | Medium | Fixed in d4ced41 (`apply_within`); gestures on the same knob are never ended, so two drags can fold into one step |+| ce64dae | Params never validated outside the UI setter: API/MCP JSON reaches the shader and the file | Low | NaN refused and tidied in d4ced41; range clamp at render still open |+| ce64dae | No transition panel on the phone | Low | Open |+| ce64dae | Tests: one export copy test, one GPU test that skips without an adapter unless `CONCAT_REQUIRE_GPU` | Low | Open |+| 64e9dcd `--software` | `--software --project X` refused (`launch.rs:76`) | Low | Open |+| 64e9dcd | Non-atomic `Preferences::save` (`prefs.rs:267-271`): a second instance can write `software_gpu=false` back | Low-Medium | Open |+| 64e9dcd | `CONCAT_GPU=software` leaks into the Linux update relaunch (`updates.rs:620-631`) | Low | Open |+| 64e9dcd | Doc drift: "software drawing" vs "safe graphics" | Low | Fixed in d4ced41 |+| 64e9dcd | No Linux `.desktop` action | Low | Open |+| f2d0637 bundled cards | Keyed by id only: a changed shader shows the old shipped card until `scripts/bundle-cards.sh` is re-run | Low | Open |+| f2d0637 | ~100 JPEG decodes on the UI thread on first shelf build, again after every reload | Low (unmeasured) | Open |+| b18c2fe / a9b2df6 FXC pin | Correct; a system-wide `WGPU_DX12_COMPILER=dxc` brings #418 back | Low | Open by design |+| be60358 shortcut reset | Correct (`input_items.rs:74-107`) | — | — |+| f121747 Concat sounds removed | `docs/AUDIO_LIBRARY.md:3-8,39-40` still describes the shelf; dead UI/Rust left (`audio-page.slint:71-80`, `studio.rs:11052`, `panes/sounds.rs:176`) | Low | Open |+| a1fd411 phone pills | A drag that starts on a pill never reaches the strip's TouchArea (`media-pane.slint:1440-1442`), so off-screen shelves stayed unreachable by touch | Medium | Fixed in d4ced41 (pills forward drags) |+| a1fd411 | Phone opened on group 0, not All | Low | Fixed in d4ced41 |+| a1fd411 | `sounds.all` reused, masculine agreement in es/pt-BR/fr/it | Low | Fixed in d4ced41 (`common.all`) |++## 2. Phone shell++- **High, Android Back quits the app**, even with a sheet open or an export+ running: nothing handles `Key.Back` (`app.slint:504-539` handles Escape+ only), and Slint finishes the activity on an unhandled Back+ (`javahelper.rs:788-811`).+- **High, a touch on a clip moves it at 4 px** (`lanes.slint:591-592`;+ Android slop is ~8 dp). Every press selects; past 4 px it is a committed+ `MoveClips` with an undo step (`studio.rs:4804-4818, 4978`). Swipe-scrub+ only starts on empty lane or the 24 px ruler (`editor.slint:726`).+- **Medium-High, pinch selects and can move clips**: the first finger's+ press already selected; Slint synthesises a release with the second+ finger's id (`input.rs:2271-2276`).+- **Medium, toast not inset**: under the nav bar / home indicator+ (`app.slint:975-981`, `toast.slint:87-88`). No soft-keyboard inset+ anywhere (Slint passes it, `javahelper.rs:814-830`; the app never reads+ it).+- **Medium, self-assigning controls unchanged**: `Select`+ (`select.slint:140`), `SegmentedControl` (`segmented-control.slint:47`),+ `Switch` (`controls.slint:561`). Concrete: Captions shows Top, Rust holds+ Bottom after a reopen (`captions.slint:226,235`, `captions.rs:409-419`).+- **Hit targets** under 44-48 px throughout: transport 36 px, Play 40 px,+ zoom chips 34, shelf pills 24, `Theme.h` 28, `Switch` 30×17; Android has+ no extra UI scale (`platform.rs:90`).+- Hold-to-repeat keeps firing after the finger slides off+ (`controls.slint:61-78`). No RTL layout for ar/fa/he/ur. No accessibility+ properties in `ui/phone`. No Slint UI tests anywhere.++## 3. Keymap, Audio page, Voiceover++- **Medium, destructive chords fire while typing**: Slint's `TextInput`+ ignores non-standard Ctrl chords (`items/text.rs:1120`), so they walk up+ to the keymap. Ctrl+B in a title field splits the title clip, Ctrl+I+ opens Import; in read-only fields Backspace deletes the selected clip.+- **Medium, no guard for open sheets**: with Settings/Export open, Space,+ Delete and Q/W act on the timeline behind (`lib.rs:1621`).+- **Medium, capture's `reserved()` unreachable**: KeyBindings match first+ (`input_items.rs:723-745`), so Ctrl+W closes the project while a+ Shortcuts row is listening; Escape closes Settings, so "Escape declines"+ is dead.+- **Medium, bindings by character, not physical key**: every letter+ shortcut is dead on Cyrillic, Greek, Hebrew, Arabic and Persian layouts+ (`winitwindowadapter.rs:1324`), and the app ships ru/uk/fa/ar/he.+- Alias conflict holes in `bind`/`reset` (`keymap.rs:257-313`).+- **Concat sounds trust anchor**: v0.2.8 still ships the "Load Concat+ sounds" button against `github.com/concatapp/assets`. Verified at audit+ time: `ConcatApp` is an organization the maintainer's account belongs+ to, so not squattable. Still: no signature, mutable `main`, digest in the+ same manifest, no download size cap (`models.rs:224-305`), Cancel never+ set (`panes/sounds.rs:113,333`), `adopt_sound` copies and probes on the+ UI thread (`panes/sounds.rs:584-637`).+- **Medium, Android microphone permission dead end**: `askMicrophone`+ returns before the result, with no `onRequestPermissionsResult`+ (`ConcatFiles.java:169-185`); a second denial has no route to settings.+- Voiceover takes: retake orphans pile up in `project/audio`; closing mid+ take drops the take from the project (`studio.rs:8414`).+- Locales: `studio.merge*` English in 21 locales, `settings.scale*` in 22.++## 4. Startup, close, project lock++- **High, File › Open with a project open dropped it unsaved and leaked+ its lock** (`start.rs:427-457` → `studio.rs:8162-8212`). Fixed in+ 2abfcf7.+- **Medium, Android saves nothing when the activity ends**: Slint ends the+ loop on `MainEvent::Destroy` and ignores Pause/Stop/SaveState+ (`androidwindowadapter.rs:271-273`); `leave(0)` follows with no+ `close_project` (`lib.rs:2433-2434`).+- **Medium, safe start treats any early death as a driver crash, for+ good** (`lib.rs:2454-2475`, `SAFE_AFTER` 15 s): a Task Manager kill, a+ logoff (winit 0.30.13 handles no `WM_ENDSESSION`), or a panic opening a+ project. Since b5e6c08 that also puts exports on WARP + FXC. The notice+ shows once. Errors returned before the loop (`lib.rs:253,259,265`) drop+ the mark instead.+- **Medium, `--smoke` leaves the start mark** (`lib.rs:2381,2385` exit+ without clearing it): the next real launch is forced into software.+- **Medium, a remembered interface scale is not applied at launch**: the+ zero-delay timer fires in `NewEvents(Init)` before the winit window+ exists (`lib.rs:2416-2419`, slint winit `event_loop.rs:213-226`), so+ `apply_ui_scale` returns early (`platform.rs:65-67`). Reasoned, not run.+- **Medium, the project lock opens unguarded on any I/O error**+ (`concat-api/src/lib.rs:183-189`): OneDrive or antivirus sharing+ violations on Windows, ENOLCK on network mounts.+- Event-loop failures exit with code 1 and no dialog (`lib.rs:2425-2431`);+ phones have no failure report and no safe start (`platform.rs:537`,+ `lib.rs:2455`).+- Low: a bare path argument (drag a folder onto the shortcut) refuses to+ start (`launch.rs:74-79`); `prefs.json` is world-readable on Unix and+ non-atomic (`prefs.rs:267-272`) though `discovery.rs:11-14` claims+ owner-only; Quit-without-saving during a version switch never starts the+ new build; closing during an export kills it without asking.diff --git a/audit/2026-10-20/window.md b/audit/2026-10-20/window.mdnew file mode 100644index 00000000..ec6c026d--- /dev/null+++ b/audit/2026-10-20/window.md@@ -0,0 +1,105 @@+## Window crate (src/crates/concat), HEAD ce64dae, 2026-10-20++Method: I read the code at HEAD and `git log`/`git show` since ab21457 (15 Oct). I ran `scripts/locales.py --check` (python only). I did not build, run cargo or open GitHub, so CI and on-device behaviour are **unverified** throughout.++**Coverage gap, stated plainly.** Four parallel deep-reviews did not report back before hand-off. They covered: today's commits (be60358, f121747, 64e9dcd, a1fd411, f2d0637, ce64dae, b18c2fe/a9b2df6); keymap/Shortcuts; the Audio page with its CC0 download shelf; Voiceover and batch volume; phone gestures (glide.rs, pinch, hold-to-repeat, Mask, Canvas and Adjustments sheets, Android back); the project lock (68ceffe); `--mcp` and the Automation token; and whether fa651ef's scale fix is correct against the Slint winit source. Those areas are **not audited** here except where a line below says I checked something myself. The rest of the crate changed by 99 commits and +21.9k/-4.6k lines in 5 days, so this pass is necessarily partial.++### Numbers (measured)+| | 15 Oct | 20 Oct | Δ |+|---|---|---|---|+| Rust lines (src/) | 24 885 | **33 163** | +8 278 (+33 %) |+| Slint lines (ui/) | 30 081 | **34 913** | +4 832 |+| studio.rs | 11 161 | **12 739** | +1 578 (38 % of the Rust) |+| `impl Studio` | 8 405 lines, 219 methods, 94 fields | **9 854 lines (2251-12104), 247 methods, 106 fields** | worse |+| `fn start` (lib.rs:196-2435) | 1 870 | **2 240** | +370 |+| Largest fns | publish_lanes 389 | publish_lanes 396, publish_chrome 353, key_editor_rows 330, stage_dragged 288, clip_set 264, menu_bar 242, settings `update` 285 | |+| unwrap/expect (all / non-test) | 59 / 4 | **112 / 4** | non-test: host.rs:174, studio.rs:2932, studio.rs:10219, panes/start.rs:230; all four are sound invariants |+| `unsafe` lines | 15 | **34** | +19, of which 20 are in the new ios.rs (objc msg_send) and 9 in wayland_drop.rs. Runtime set_var sites: lib.rs:2472, platform.rs:427 |+| `#[test]` | 96 in 22 files | **155 in 28 files** | +59 |+| Locales | 26 files (+stray vn), 673 keys, --check exit 1 | **25 files, 881 keys, --check exit 0**, 707 values still English | |+| Studio splits | consolidate.rs | + backdrop.rs 306, masking.rs 409 | |++### Re-verified findings from 15 Oct+| # | Finding | Status at HEAD | Evidence |+|---|---|---|---|+| 2 | iPhone Import does nothing | **FIXED in code** (device unverified) | lib.rs:200-201 `ios::install()`; ios.rs:100-133 installs the picker and publisher |+| 5 | >24 titles drop out mid-drag | UNCHANGED | studio.rs:2953-2970 hold_still per title; concat-media/src/pool.rs:541 STILLS_HELD=24 |+| 6 | Shared save scratch | **FIXED** | concat-host/src/projects.rs:300-330 per-call `create_new` temp (pid + nonce). The doc at :267-271 still says "a project folder has none [lock] today", which is stale since 68ceffe (Low drift) |+| 8 | Captions sheet self-assigning bindings | UNCHANGED | select.slint:140, segmented-control.slint:47 assign `root.current`; captions.slint:193, 204, 226, 235 bind one-way |+| 9 | studio.rs growth | **WORSE** | 12 739 lines (+1 578); two small splits do not offset it |+| 12 | A phone that can't start vanishes; iPhone files hidden | PARTLY | Files: plist keys added (src/scripts/ios-app.sh:99-103). Phone startup failure is still a no-op (platform.rs:536-537). Android gets no safe start (lib.rs:2455-2457) |+| 13 | locales.py passes with missing keys | UNCHANGED | scripts/locales.py:193-194 prints "missing" but never sets `failed` |+| 17 | `--smoke` stops before the window exists | UNCHANGED | lib.rs:2366-2385 exits before `app.run()` (:2417), so the swapchain and first frame (#389-class crashes, 369fdfa) are never exercised |+| 18 | Window smalls | PARTLY | meters-pane.slint and METERS_MIN_W are gone with the VU rewrite. host.rs:102, 281, 359, 386, 406 still drop `invoke_from_event_loop` errors |+| 22 | 4 px drag threshold below finger slop | UNCHANGED | timeline/lanes.slint:591-592. The new finger gesture uses 6 px (:1204-1206), still under Android's 8 dp |+| 23 | #151 returns after Browse | UNCHANGED | panes/start.rs:255 and :295 set the location without recomputing the free "Untitled" name (only Compose at :243 does) |+| 24 | Detach unmutes deliberately muted video | UNCHANGED | studio.rs:12124-12134 (`Again` for any muted video), :11953-11964 |+| 25 | Unknown argument refuses to start | UNCHANGED | launch.rs:73-78. Windows "Open with" and Linux `%U` still produce an error box |+| 26 | Commits that close issues | **FIXED this period** | none of the 179 commits since 15 Oct uses a closing keyword (word or URL form). PR bodies of the merged #377/#401/#402/#403 are unverified |+| 27 | RTL laid out LTR | UNCHANGED | no layout-direction handling anywhere in ui/ or src/ |+| 33 | Doc drift; gpu.rs drops errors | UNCHANGED + new | gpu.rs:87, 122 `.ok()?` drops adapter and device errors unlogged. New drift: launch.rs:18 says "Concat (software drawing)" but the installer says "Concat (safe graphics)"; lib.rs:2469-2470 SAFETY cites "the scale above", which no longer sets env; studio.rs:10212-10215 has two stacked doc comments |+| 11th#3 | Paste/duplicate keeps shape | **WORSE** | Titles and shapes still carry only style, duration, offset_y, name (studio.rs:8007-8031). `copy_commands` (studio.rs:2106-2180) now also misses the **new** `mask`, `stabilize` and `background`, plus `cutout`, `audio_stream`, `enhanced`. Clipboard survives close (see N1) |+| 11th | Literals bypass i18n | UNCHANGED + new | studio.rs:1140 "New title", :1225 "Other", :5352 "Title", :11204 "Freeze frame", :4097, :7820, :8328-8329 English notices; consolidate.rs:52, :107 pass `error.to_string()` |+| 12th | platform.rs SAFETY "single-threaded" untrue | UNCHANGED | platform.rs:424-427 sets SLINT_WGPU_CPU *after* `Gpu::acquire()` (:409) has created a wgpu device. On lavapipe that device has started threads, so "single-threaded" is false |+| H1 | main red; stray vn.json | **FIXED locally** | vn.json deleted (d3f612e); `locales.py --check` exit 0. CI state unverified |+| H2 | Phones never built | CONFIG FIXED, runs unverified | .github/workflows/mobile.yml:77-84 now builds on push to main (5195bb2) |+| M1 | Interface scale replaces display scale | **FIXED per code** (correctness vs Slint internals unverified) | platform.rs:44-102 multiplies winit's `scale_factor()` and re-applies after a monitor change; SLINT_SCALE_FACTOR set_var removed |+| M2 | Export picker shows a codec it won't use | PARTLY | Details now read "not on this device" (dialogs/export.slint:329-332). But the Select still gets no `disabled` list, so select.slint:140 assigns AV1, panes/export.rs:149-153 refuses it, and the picker keeps showing AV1 while the export writes H.264. One-line fix: pass `disabled` |+| M3 | Phone dialogs/toasts ignore safe area | PARTLY | Every sheet and floating panel now takes the insets (app.slint:642-645, 683-686 … 927-928; modal.slint:46-99). **The toast does not**: app.slint:975-981, toast.slint:86-88 put it 16 px off the window bottom, i.e. under the Android nav bar or iPhone home indicator |+| M4 | Duplicate is two undo steps | UNCHANGED | studio.rs:7924 `place_copy(..., None)`; AddClip of the whole file (:8063-8071), then a separate Batch (:8072-8073) |+| M5 | Merge re-encodes instead of a lossless join | UNCHANGED | studio/consolidate.rs untouched since; the window never issues `MergeClips` (concat-project commands/mod.rs:623) |+| L1 | Runtime set_var with false SAFETY | PARTLY | Scale set_var removed. safe_start's set_var (lib.rs:2472) is pre-thread and plausibly true; platform.rs:427 is not (above) |+| L2 | merge_blocked in every lanes publish | UNCHANGED | studio.rs:9141 inside publish_lanes (8897-9292) |+| L3 | Progress bar redraws every frame | UNCHANGED | primitives/modal.slint:416, 455 `animation-tick()` |+| L4 | "Apply to all" silent and destructive | MOOT | feature removed in 13f6d4f (-286 lines) a day after it merged |+| L5 | Contributor strings English in every locale | PARTLY | merge (10 keys) and interface scale (3 keys) still English in 21-22 of 24 locales (`--check --same`); id and pt-BR translated |+| L6 | vn.json edited while never compiled in | FIXED | deleted |+| L7 | ClearType never seen on Windows | **MATERIALISED, then fixed** | 369fdfa: bdcaedf crashed Windows on the first frame of text (MSVC by-value return ABI in `top_props`). Fixed before v0.2.8. Code lives in src/patches/i-slint-renderer-skia (outside this crate). SystemParametersInfoW still runs per frame (dx12.rs:75-78) |+| L8 | Captions cancel/restart collide | UNCHANGED, worse than suspected (reasoned) | panes/captions.rs:184-188 Cancel clears `running`, so Begin can start a second run. The first run's late `Finished(Err cancel)` then runs `self.subject.take()` (:192), stealing the **second** run's subject. The second run's Ok then returns at :199-201 and its captions are dropped silently; :229 also closes the sheet. Cancellation is detected by `error.contains("cancel")` on English text |+| L9 | Merge interrupted by quit leaves media/merged-* | UNCHANGED | lib.rs:2511-2524 `_exit`/`process::exit` skip Drop guards |++Tally: 6 fixed, 8 partly, 15 unchanged, 3 worse, 1 moot.++### New findings (verified unless marked)+**MEDIUM**+- **N1 Paste after switching projects inserts the wrong footage.** `release_project` (studio.rs:8370-8432) never clears `self.clipboard` (field :860, only set at :7959). Ids are per-project counters (`IdMint`, concat-project/src/commands/mod.rs:890-900: "m1", "m2" …), so a copied clip's `media_id` names a *different* file in the next project. Scenario: copy a clip of `m3` (beach.mp4) in project A, close, open project B, Ctrl+V. `place_copy` (studio.rs:8063-8071) adds B's `m3` (say interview.mov) with A's trims, keys and effects. If B has no `m3`, the paste fails with nothing said. Paste stays enabled (:11169). Fix: clear the clipboard in `release_project`, or key it by project.+- **N2 Copy/Paste/Duplicate drops the features added this week.** `copy_commands` (studio.rs:2161-2179) builds its `ClipPatch` without `mask` (16a3d13, #239), `background` (Canvas, 244765e) or `SetClipStabilize` (915babd), and also skips `cutout`/`audio_stream`/`enhanced`. Model fields: concat-project model.rs `Clip { … cutout, mask, stabilize, … background }`. Scenario: mask a clip with a heart, Duplicate: the copy is unmasked and unstabilized. The only test (studio.rs:12266) covers split frames. Add a field-exhaustiveness test (destructure `Clip` without `..`).+- **N3 One early death of any cause puts the whole app on the CPU rasteriser for good.** safe_start (lib.rs:2454-2475) flips `software_gpu=true` permanently when the start mark is left. Things that leave it: any death before 15 s (SAFE_AFTER, :2437), including Task Manager, a logoff or shutdown, a panic opening a corrupt `--project`, or a blur stall killed by hand. From then on `CONCAT_GPU=software` is set every launch, and the shared device is WARP/lavapipe for preview, effects *and export* (gpu.rs:76-87; Host::start(gpu) lib.rs:256). You get one toast (lib.rs:282-284) and a switch in Settings. Nothing re-tries the GPU, even after a driver update. Also: `--smoke` exits via `leave()` (lib.rs:2381, :2385) without dropping `startup_mark`, so any smoke run on a user machine forces the next real launch into software. Design trade-off defensible for #339/#389, but it should be one-shot (retry the GPU after N clean software launches, or ask).+- **N4 Shortcut, scale, GPU and every other preference can vanish on a crash.** prefs.rs:267-271 `fs::write` straight over the file (not atomic). prefs.rs:258-262 then treats any parse failure as all-defaults. A crash or power loss mid-save (saves happen on every settings toggle, including safe_start's own save at :2467) silently resets rebound shortcuts, interface scale, theme and software_gpu. The project save got atomic temp+rename this week; preferences did not.++**LOW**+- N5 The startup Tripwire is shared by concurrent windows (allowed, lib.rs:153-155). Window 2 overwrites window 1's mark, and window 1's 15 s timer (lib.rs:2408-2412) deletes window 2's, so a crash of window 2 in its first seconds is not recorded. Windows pid reuse can also make `alive()` lie. (reasoned)+- N6 A toast on phones sits under the system bars; see M3 (app.slint:975-981).+- N7 Captions finish matching on English error text (`error.contains("cancel")`, panes/captions.rs:229). Also see L8.+- N8 Doc drift: launch.rs:18 vs the installer name; projects.rs:267-271 "no lock"; lib.rs:2469-2470 SAFETY cites a removed set_var; studio.rs:10212-10215 stacked docs.+- N9 English literals added or kept in user-facing notices (studio.rs:4097, 7820, 8328-8329).++**Not audited (agent reviews missing; treat as open questions)**: today's commits (shortcut-row reset, Audio page shelf, `--software` installer entries, phone shelf pills, the 100+ bundled cards, transition knobs and their undo granularity); keymap layout handling (char vs physical key, AZERTY), Q/W/E bindings while typing; the CC0 sound download (URL, integrity, path traversal); Voiceover mic-denied paths; glide/pinch/hold-repeat timers on Android; Android back button; the project lock on removable or network drives and after a kill; `--mcp`/Automation token strength and bind address; whether `dispatch_event(ScaleFactorChanged)` survives Slint's own resize handling.++### Commits since 15 Oct, short verdicts (only those I read)+- d3f612e vn.json removed: fixes H1.+- e75af9b/1b7e8e6/9197340 saves: per-save temp + ticket ordering (projects.rs:255-330), and close refuses on a failed save then offers quit-without-saving (studio.rs:8345-8368). Sound, and lib.rs has two tests (lib.rs:2599, :2677).+- fa651ef scale: right approach (platform.rs:44-102); runtime correctness unverified.+- b5e6c08/e09d154/64e9dcd safe graphics: works as designed; sticky (N3); smoke leaks the mark.+- 8dfa7bb iOS import: wired; never run here.+- 13f6d4f removed a68641c's contributor feature a day after merging it: churn, not a fix.+- 16a3d13 mask / 915babd stabilize / 244765e canvas: none of them reach copy/duplicate (N2).++### Scores (window crate)+| Aspect | 15 Oct | 20 Oct | Why |+|---|---|---|---|+| Code quality | 6 | 6 | Careful local code, 4 sound non-test expects in 33k. Against: copy paths forgot three new clip fields; clipboard leaks across projects; false SAFETY persists |+| Design architecture | 7 | 6 ↓ | The message/pane split holds, but every feature still lands in `Studio` (106 fields, 247 methods) and `start()` grew to 2 240 lines |+| Design philosophy | 6 | 6 | Platform-native intent is visible (iOS pickers, phone sheets, safe area). Against: features merged then deleted next day; a sticky safe mode instead of a measured one |+| Maintainability | 4 | 3 ↓ | +8.3k Rust in 5 days; studio.rs 12.7k; publish_lanes 396, publish_chrome 353; two 300-400-line splits are cosmetic against that |+| User likeability | — (crate) | 6 | Real gains: iOS import, sheets inside phone bars, home page, shortcuts, VU, voiceover, masks, scale fixed. Losses: wrong-footage paste, copies losing masks, a WARP-forever trap after one bad start, toasts under the nav bar |+| Performance | 6 | 6 | Unchanged hot spots (merge checks per publish, per-frame progress tick); meters tick at 33 ms and stop when quiet (meters.rs:26). Software mode can stick on CPU rendering (N3) |+| Scalability | 5 | 5 | 24-still cap unchanged; per-project id counters make cross-project anything unsafe |+| Security | 6 | 6 (partly unverified) | Save temp uses `create_new` against planted symlinks. Network shelf, MCP and token not reviewed here |+| Testing | 5 | 6 ↑ | +59 tests (155 in 28 files), incl. close/save races and the lock (per commit; not re-read). Still none on copy-field completeness, startup marks or window runtime; smoke never reaches the first frame |+| Documentation | — | 5 | Dense, honest module docs, but drift keeps accumulating (N8) and the docs describe intent more confidently than the code earns |++### Top three to do+1. Clear the clipboard in `release_project` and make copy field-exhaustive (N1, N2). Small, and wrong-footage paste is a correctness bug users will hit.+2. Make safe start one-shot or self-healing, and drop the mark on the smoke path (N3). Write prefs atomically (N4).+3. Pass `disabled` to the export codec Select (M2), inset the toast (M3), and put a run id on captions (L8). Each is a few lines.diff --git a/audit/AUDIT-2026-10-20-QUICK.md b/audit/AUDIT-2026-10-20-QUICK.mdnew file mode 100644index 00000000..ccc52109--- /dev/null+++ b/audit/AUDIT-2026-10-20-QUICK.md@@ -0,0 +1,91 @@+# Concat audit, 20 October 2026: the quick version++The one-screen scorecard for [`AUDIT-2026-10-20.md`](AUDIT-2026-10-20.md).+Code at `c41865e`, 175 commits past the 15 October audit, with v0.2.8+published today (tag `79116bd`). The reviewers' own reports are in+[`2026-10-20/`](2026-10-20/).++## In one paragraph++0.2.8 is the biggest release yet and the first whose notes are honest+about what was not tested. About 205 000 people have downloaded a Concat+build, 78% of them on Windows or Android. The engine is in better shape+than the shell: the server work two audits asked for landed, phones now+decode on MediaCodec, and 984 tests pass with clippy clean. But the things+that lose a user in the first ten minutes are still there. A single early+crash puts the app on the software renderer for good. Android's Back+button quits mid-export. A tap on a clip moves it. Windows is unsigned and+in no store. CI on `main` has been red or cancelled for most of a week, so+it tells nobody anything. And none of today's fixes are in any release.++## Scores++| Aspect | 15 Oct | Today |+|---|:-:|:-:|+| Code quality | 6 | 6 |+| Design architecture | 7 | 7 |+| Design philosophy | 5 | 6 ↑ |+| Maintainability | 5 | 4 ↓ |+| User likeability | 6 | 6 |+| Performance | 5 | 6 ↑ |+| Scalability | 5 | 5 |+| Security | 5 | 5 |+| Testing | 5 | 6 ↑ |+| Documentation | 6 | 5 ↓ |+| CI and release | 4 | 5 ↑ |+| Distribution and trust (new) | — | 3 |++## The numbers++| | Today | Δ since 15 Oct |+|---|---:|---:|+| Rust lines, every `.rs` under `src/crates` | 121 054 | method changed, not comparable |+| Slint lines | 34 913 | +4 832 |+| `#[test]` functions | 988 | +210 |+| `cargo test --workspace` | 984 passed, 0 failed, 3 ignored | was 773/1/1 |+| clippy `-D warnings`, fmt | pass | |+| `locales.py --check` | pass (881 keys; 707 values still English) | was failing |+| `unsafe` (grep) | 138 | +28 |+| `studio.rs` | 12 739 lines | +1 578 |+| `start()` | 2 240 lines | +370 |+| Effect packages | 127 (102 picture, 25 audio) | |+| CI on `main` | red or cancelled most of the week; last two pushes running | |+| Open issues / PRs | 141 / 16 | +35 / +10 |+| Downloads, all releases | ~205 000 | v0.2.7 alone 126 218 |++## Fixed today (on `main`, in no release yet)++#418 Windows start crash from a stray `dxcompiler.dll` (b18c2fe, a9b2df6) ·+garbled-window escape, "Concat (safe graphics)" (64e9dcd) · phones reach+every transition and filter shelf (#415, a1fd411, d4ced41) · effect cards+ship inside the app (#416, f2d0637) · transition knobs settable (ce64dae,+d4ced41) · File › Open no longer drops the open project unsaved and leaks+its lock (2abfcf7) · Android engine library aligned for 16 KB pages, with+a CI check (c41865e) · shortcut reset (be60358) · private vulnerability+reporting switched on (it was off while SECURITY.md pointed at it).++## Open, High++| # | What | Where |+|---|---|---|+| H1 | Any death in the first 15 s puts every later launch on WARP/lavapipe, export included, for good; `--smoke` leaves the mark too | `concat/src/lib.rs:2454-2475, 2381` |+| H2 | Android Back quits the app with a sheet or an export open; nothing is saved when the activity ends | `app.slint:504-539`, `lib.rs:2433` |+| H3 | Phone: a 4 px wobble on a clip moves it with an undo step; a pinch selects and can move; targets 24-36 px | `lanes.slint:591`, `studio.rs:4804-4978` |+| H4 | CI on `main` is no signal: phone runs cancelled 50 of 60 times, lint steps abort before tests | `mobile.yml:92`, `ci.yml:136-141` |+| H5 | Unsigned on Windows, ad-hoc on macOS, in no store; the updater trusts a digest from the release it checks | `build-app.yml:1036`, `updates.rs` |+| H6 | macOS bundles say 12.0; CI shows linked libraries needing 14 or 15 and only warns | `build-app.yml:96, 294-302` |+| H7 | Export recovery opens a second high-performance DX12 device, the road #202 says crashes hybrid NVIDIA laptops | `concat-export/src/lib.rs:1530` |+| H8 | Captions start early (#438): no voice detection, no token timestamps | `concat-speech/src/transcribe.rs:481-545` |++## Do first++1. Ship today's fixes: #418 and the 16 KB alignment are start-up failures in the wild.+2. Point the website at 0.2.8 and fix its checksum link (#458).+3. Safe start: trip only on a crash, retry the GPU, clear the mark on `--smoke`.+4. Android Back closes the top sheet; save on pause and destroy.+5. Phone timeline: long-press to move, 8 dp slop, 44 px targets.+6. CI: lint in its own job, no cancelling on `main`, phone runs required.+7. Windows signing or the Microsoft Store; Android developer registration.+8. An honest macOS floor, and a failing minimum-OS check.+9. Land D3monX's caption timing PRs once they're up.+10. Merge the four approved PRs (#445, #446, #441, #424): blocked on a permission only the owner can grant.diff --git a/audit/AUDIT-2026-10-20.md b/audit/AUDIT-2026-10-20.mdnew file mode 100644index 00000000..d580560f--- /dev/null+++ b/audit/AUDIT-2026-10-20.md@@ -0,0 +1,287 @@+# Concat audit, 20 October 2026++Code at `c41865e` on `main`, 175 commits past the [15 October+audit](AUDIT-2026-10-15.md). v0.2.8 was published today from tag+`79116bd`. The one-screen version is+[`AUDIT-2026-10-20-QUICK.md`](AUDIT-2026-10-20-QUICK.md).++Six 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`](2026-10-20/window.md) | The `concat` crate: shell, settings, dialogs, copy and paste |+| [`window-subreviews.md`](2026-10-20/window-subreviews.md) | Today's commits, the phone shell, keymap/audio/voiceover, startup/close/lock |+| [`host-server-api-cli.md`](2026-10-20/host-server-api-cli.md) | concat-host, concat-server, concat-api, concat-mcp |+| [`media-render-export-core.md`](2026-10-20/media-render-export-core.md) | concat-media, concat-render, concat-export, concat-core |+| [`project-effects-ml-phones-ci.md`](2026-10-20/project-effects-ml-phones-ci.md) | concat-project, concat-effects, speech, vision, phone shells, CI and release |+| [`product-distribution.md`](2026-10-20/product-distribution.md) | Releases, reach, stores, signing, GitHub state, README claims |++"Verified" below means traced in code, in a shipped artifact or in CI logs.+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 is real.** About 205 000 binary downloads across releases. v0.2.7+drew 126 218 in its week as Latest: Windows 43.1%, Android 35.0%, iOS 8.1%,+macOS 7.3%, Linux 6.5%. Traffic is mass-market (Google, Facebook), not+developers. Retention is not: only about 1% of v0.2.7 installs came back+through the in-app updater (the manifest was fetched 1 339 times).++**The engine is the strongest part.** The server work that slipped through+two audits landed with tests (363800f): no pre-auth fd exhaustion, no+blocking stop, a real connection count. Phones decode 8-bit H.264, HEVC,+VP9 and AV1 on MediaCodec (2969e02). Packets refused on a full codec are+sent again (ac8a654). The workspace runs 984 tests with 0 failures, and+clippy and fmt are clean.++**The shell is where users are lost.** The first-ten-minutes failures from+the last two audits are still mostly open, on the two biggest platforms:+safe start that never lets go, Android Back that quits, a phone timeline+that moves clips on a tap, an unsigned Windows build in no store. Feature+work keeps outrunning them: 175 commits in five days, `studio.rs` up+1 578 lines to 12 739, `start()` up 370 to 2 240.++**Nothing fixed today has reached anyone.** v0.2.8 was tagged before+today's work. The website still links v0.2.7 and its checksum link now+points at 0.2.8's file, so a site download fails its own hash check.++## 2. Scores++| Aspect | 15 Oct | Today | Why |+|---|:-:|:-:|---|+| Code quality | 6 | 6 | Careful local code and few non-test unwraps (4 in the window crate), against copy paths that drop new clip fields and a clipboard that crosses projects |+| Design architecture | 7 | 7 | The server, MediaCodec and lock designs are sound; `Studio` and `start()` keep absorbing every feature |+| Design philosophy | 5 | 6 ↑ | Platform-native moves (iOS pickers, phone sheets, MediaCodec), honest release notes; against a safe mode that never lets go |+| Maintainability | 5 | 4 ↓ | +8.3k Rust lines in the window crate in five days; export `lib.rs` 3 770 lines, render `gpu.rs` 3 695 |+| User likeability | 6 | 6 | Masks, stabilization, voiceover, audio page, CapCut keys, against start-up traps, phone touch and 65% of outside issues unanswered |+| Performance | 5 | 6 ↑ | Phone decode on the chip; offset by 3 MediaCodec slots never reclaimed, WARP+FXC stuck exports, a blurred background running the effect chain twice |+| Scalability | 5 | 5 | Unchanged caps (24 held stills, 512 MB reader pool on any device); per-project ids make cross-project paste unsafe |+| Security | 5 | 5 | Server hardened; still an unsigned update chain, an auth throttle anyone local can trip, MCP hints that understate writes, world-readable prefs on Unix |+| Testing | 5 | 6 ↑ | +210 tests; still no Slint UI tests, no device test that runs, nothing on safe start, copy completeness or Android lifecycle |+| Documentation | 6 | 5 ↓ | Honest notes, but a stale website, a false keybind line in the 0.2.8 notes, README platform ticks ahead of reality, doc drift in a dozen places |+| CI and release | 4 | 5 ↑ | A release passed all 16 jobs honestly; `main` is no signal and a release still passes unsigned |+| Distribution and trust | — | 3 | GitHub and a stale site only; unsigned or ad-hoc on three OSes; no Android developer registration; a name squatter org |++## 3. Today's work, reviewed like anyone else's++The owner asked for strict scrutiny of every input, and that includes the+work done in this session. A dedicated review read every commit from today.+What it found, and what happened:++| Commit | Finding | Now |+|---|---|---|+| ce64dae transition knobs | Rows rebuilt on every publish: a slider drag stopped after one step (High) | Fixed d4ced41 |+| ce64dae | One undo step per tick (Medium) | Fixed d4ced41; two drags of one knob can still fold into one step |+| ce64dae | NaN knobs could save as null and make the project unreadable | Fixed d4ced41 |+| a1fd411 phone pills | A drag starting on a pill never scrolled the strip, so off-screen shelves stayed unreachable: #415 half-fixed (Medium) | Fixed d4ced41 |+| a1fd411 | Phone opened on the first shelf, "All" in the wrong grammatical gender | Fixed d4ced41 |+| f121747 Concat sounds | Left eight unused strings; the locale check went red on CI and hid clippy and tests behind it | Fixed baaeb2a |+| 64e9dcd safe graphics | Doc drift on the entry's name | Fixed d4ced41 |+| Labels | `testing-requested` put on #415, #416, #418 whose fixes are in no release | Removed |+| f2d0637 bundled cards | A changed shader keeps showing the old shipped card until re-bundled; ~100 decodes on the UI thread | Open, Low |+| 64e9dcd | Non-atomic prefs can lose the safe-graphics setting to a second instance | Open, part of M-prefs below |++Two High findings from the reviews were fixed during the audit:++- **File › Open with a project open dropped it unsaved and kept its lock**+ (`start.rs:427-457` → `studio.rs:8162-8212`). Fixed in 2abfcf7: the open+ project is saved and released first; a failed save refuses the switch.+- **The shipped Android engine library is aligned for 4 KB pages.** Checked+ on the v0.2.8 APK: `libconcat_android.so` LOAD align 4096 beside three+ 16384 libraries. A 16 KB-page phone refuses it; Play requires it. Fixed+ in c41865e with a CI check that fails on the shipped APK.++## 4. Carried findings from 15 October++Across the four area reports, roughly a quarter of the earlier findings are+fixed, a third partly, and the rest unchanged; a handful are worse. The+per-item tables are in each area report. The ones that matter most:++| 15 Oct | Then | Now |+|---|---|---|+| H1 `main` red, tests hidden behind a lint step | High | Recurred this week; still the same job layout |+| H2 Windows unsigned, signing step exits 0 | High | Unchanged |+| H3 iOS without Import | High | Fixed in code (8dfa7bb), untested on a device; #294 iOS launch crash unanswered 11 days |+| H4 Android launch check never runs the app | High | Partly: the x86_64 build is started; the shipped arm64 APK never is |+| H5 Android recreation on config change | High | Fixed: all 13 flags in the shipped manifest |+| H6 Intel Mac on macOS 12 | High | Partly: ONNX Runtime loads lazily, but CI shows other libraries needing 14/15 (new H6 below) |+| N1 Android CPU decode | High | Partly: 8-bit on MediaCodec; 10-bit HEVC still CPU; 3 slots never reclaimed |+| N2 Chained transitions climb a lane | High | Unchanged, and a caption there is now forced opaque |+| M1 Interface scale replaced the display's | Medium | Fixed in Settings; a remembered scale is not applied at launch (reasoned) |+| M6 Remote API fd exhaustion | Medium | Fixed (363800f) |+| M16 README promised gRPC and a CLI | Medium | Fixed |++## 5. Open findings, High++**H1. Safe start makes any early death permanent.** `lib.rs:2454-2475`+treats a death in the first 15 s as a driver crash and sets+`software_gpu` for good. A Task Manager kill, a Windows logoff (winit+0.30.13 handles no `WM_ENDSESSION`), or a panic opening a corrupt project+all count. Since b5e6c08 that also puts exports on WARP with FXC, where+the repo's own test note says package compiles "ran past five minutes".+Compiles have no deadline and Cancel cannot reach them, so a blurred+export looks hung. The notice shows once. `--smoke` exits without clearing+the mark (`lib.rs:2381, 2385`), so a developer's smoke run forces the next+launch into software. Errors returned before the loop drop the mark, so a+reported GPU failure repeats with no fallback. Sources: window N3,+media H1, startup review.++**H2. Android Back quits, and Android saves nothing on the way out.** No+`Key.Back` handler exists (`app.slint:504-539`); Slint finishes the+activity on an unhandled Back (`javahelper.rs:788-811`). Slint ends the+loop on `Destroy` and ignores Pause and Stop+(`androidwindowadapter.rs:271-273`), then `leave(0)` exits without+`close_project` (`lib.rs:2433`). Back with Export open kills the export;+a background kill loses the autosave window.++**H3. The phone timeline moves clips on a tap.** The drag threshold is+4 px against Android's ~8 dp slop (`lanes.slint:591`). Every press+selects; past 4 px it commits a move with an undo step+(`studio.rs:4804-4978`). Swipe-to-scrub only starts on empty lane or the+24 px ruler. A pinch's first finger selects clips and can move them.+Targets are 24-36 px throughout (transport, zoom chips, pills, `Theme.h`+28 px), with no phone UI scale. CapCut gates moves behind a long-press.++**H4. CI on `main` tells nobody anything.** Of 60 phone runs on `main`+since 16 October, 50 were cancelled by a newer push and 7 failed; of 60 CI+runs, 28 failed and 22 were cancelled. The locale and fmt gates still sit+before clippy and tests in one job (`ci.yml:136-141`), so a string slip+hides everything (it happened today). Releases pass only because they+re-run CI on the tag.++**H5. No trust chain end to end.** Windows unsigned (`WINDOWS_CERTIFICATE_PFX`+unset; a release still passes without it, `build-app.yml:1036`). macOS+ad-hoc and not notarised; iOS ad-hoc. No store or package manager the+project runs. The updater checks downloads against a digest published in+the same release (S1, `updates.rs`, unchanged since 12 October). The+Android signing key changed between v0.2.5 and v0.2.6, so the 15 206+people on 0.2.5 cannot update in place. No Android developer registration+ahead of Google's verification rollout.++**H6. macOS minimum version is not what ships.** The bundles declare+12.0; CI logs list ten Homebrew FFmpeg/OpenSSL libraries built for 14.0+(Intel) or 15.0 (Apple silicon), and the check only warns. The 0.2.8 note+"Intel Macs open on macOS 12 again" is unproven. (Reviewer-verified from+CI logs; not re-checked here.)++**H7. Export recovery takes the road that crashes hybrid laptops.** When+a device dies under an export on the window's lent device, recovery opens+a fresh high-performance DX12 device (`concat-export/src/lib.rs:1530`),+the second-device path #202 ties to NVIDIA hybrid crashes; the sibling+also shares the window's error counter (`gpu.rs:960-963`).++**H8. Captions start early (#438).** Whisper runs with no voice-activity+detection or token timestamps and raw segment times+(`transcribe.rs:481-545`), so each pause is swallowed into the next+caption. A contributor (D3monX) has two working branches and offered PRs;+left to them deliberately.++## 6. Open findings, Medium++Grouped; full evidence in the area reports.++- **Data and projects.** Paste after switching projects inserts the other+ project's footage under a reused media id, or fails silently+ (`studio.rs:8370-8432`, `commands/mod.rs:890`). Copy/paste/duplicate drop+ mask, stabilize, background, cutout and enhance (`studio.rs:2161-2179`).+ Preferences are written non-atomically and any parse failure resets all+ of them (`prefs.rs:258-272`). The project lock opens unguarded on any I/O+ error, including OneDrive or antivirus sharing violations+ (`concat-api/src/lib.rs:183-189`). Agent-opened projects stay locked+ until the server ends (`json.rs:332-335`).+- **Server and agents.** One shared auth-throttle bucket for every local+ caller: a web page's fetch to port 7420 keeps `concat --mcp` out+ (`json.rs:93-97, 228-240`). MCP hints call `preview_frame` read-only+ while it can overwrite any file under home (`tools.rs:445-453`). The+ shipped `concat --mcp` is never run in CI, and its panic nets do nothing+ under `panic = abort`.+- **Render and media.** Any live adjustment layer turns every packaged+ transition into a plain dissolve (`concat-export/src/lib.rs:1958-1964`):+ the most likely code-level cause of #448, which today's GPU test+ bypasses. GLES phones end an export after two 2.15 s waits. The default+ export moved to single-pass ABR at upload bitrates with x264 `veryfast`+ (~8 Mb/s at 1080p30, half the old CRF; unmeasured). 3 MediaCodec slots,+ first-come, never reclaimed.+- **Keyboard.** Destructive chords fire while typing (Ctrl+B splits the+ title being edited; Backspace in a read-only field deletes a clip); no+ guard while a sheet is open; letter shortcuts dead on Cyrillic, Greek,+ Hebrew, Arabic and Persian layouts, all of which ship as languages.+- **Phone.** Toast and tooltips not inset; no soft-keyboard inset; Select,+ SegmentedControl and Switch assign their own bindings, so Captions shows+ Top and places at the bottom after a reopen; no RTL; no accessibility+ names; microphone permission dead end on Android.+- **Downloads.** No size cap on model or sound downloads; phone FFmpeg and+ sherpa-onnx builds pulled unverified into the signed APK; the model+ mirror (`models-v1`) returns 404 and its workflow never ran.+- **Product claims.** README ticks iOS, Android and Windows ✅ against an+ unanswered iOS launch crash, an APK nobody has started and ten open+ Windows crash issues. "4K scrubs smoothly" has no measurement. The 0.2.8+ notes say "Q and E cut the selected clip"; Q and W delete left/right and+ E is unbound.++## 7. GitHub and community++- 141 open issues, 16 open PRs. 70 of 108 open issues filed by others (65%)+ have no maintainer reply; the oldest unanswered crash is #294 (iOS).+- Every open issue is labelled as of today (ten area labels added). 29+ carry `testing-requested` for fixes that are in v0.2.8; reporters are not+ notified by a label, so each needs a comment.+- Reviewed today: #445, #446, #441, #424 approved; #420 and #421 approved+ in logic, waiting on CI that has not run on the forks; #360 and #383+ changes requested (a dead `vn.json`; a model licence the package itself+ contradicts); #447 and #419 given direction. Merging is blocked by this+ session's permission settings.+- #157 (geometric masks, 31 days) was never reviewed and was superseded by+ in-house masks. #300, #187 and #275 were closed without a word.+- Flags, not acted on: the GitHub org `Concat-video-editor` (created+ 16 October) is SEO spam using the name and hotlinking the site's image.+ The owner's #363 (reuse assets from an installed CapCut) is a legal+ liability and contradicts D#18. No prompt injection found.++## 8. Competitive position++| | Concat | CapCut | Drift | OpenCut |+|---|---|---|---|---|+| Platforms | Win, Mac, Linux, Android, iOS (sideload) | Win, Mac, Android, iOS, web | Win, Mac, Linux, Android | Web; native rewrite under way |+| Distribution | GitHub releases only | Every store | Flathub (16 880 installs), Microsoft Store, Homebrew, Play | Web |+| Price | Free, no account | Pro $179.99/yr; Pro assets block export | Free | Free |+| Effects | 102 shader packages; no assets, masks, motion or beat inputs | Engine scenes: GLES passes, Lua/JS, PNG sequences, ML graphs | Qt/GLSL, 3D, add-on market | Basic |+| Agent API | MCP + JSON-RPC | None public | MCP since August | None |++Concat leads Drift on GitHub (about 205k downloads to 95k, 9.6k stars to+2k) and trails it on every store. The effect gap to CapCut is analysed in+[`reports/CapCut effects versus Concat effects.md`](../reports/CapCut%20effects%20versus%20Concat%20effects.md):+it is mostly motion, timing, masks and assets, not shader quality.++## 9. What to do, in order++1. **Ship today's fixes.** #418 and the 16 KB alignment are start-up+ failures on real machines; #416, #415 and File › Open are visible+ daily. Nothing today reaches a user until a release does.+2. **Website.** Point both domains at 0.2.8 and fix the checksum link+ (#458).+3. **Safe start (H1).** Trip only on a crash signature, retry the GPU after+ a clean software run, clear the mark on `--smoke`, show the state in+ Settings.+4. **Android lifecycle (H2).** Back closes the top sheet and asks before+ leaving an export; save on pause and destroy.+5. **Phone timeline (H3).** Long-press to move a clip, 8 dp slop, a pinch+ that never selects, 44 px targets, inset the toast.+6. **CI (H4).** Lint and locale checks in their own job; no cancelling on+ `main`; the phone workflow and the 16 KB check required.+7. **Trust (H5).** Windows signing or the Microsoft Store (registration is+ free for individuals); Android developer registration and IzzyOnDroid;+ a signed update manifest.+8. **macOS floor (H6).** Declare what the libraries need, or rebuild them+ for 12, and make the check fail.+9. **Data safety.** Atomic prefs, clipboard cleared per project, copy that+ carries every clip field, a lock that refuses on I/O errors.+10. **Server.** Per-connection throttle or none for minted tokens; release+ agent-opened projects on disconnect; honest MCP hints.+11. **Keyboard.** Guard the keymap while a text field or a sheet has focus.+12. **Captions (H8).** Review and land D3monX's PRs when they arrive.+13. **PRs.** Merge #445, #446, #441, #424 (needs the owner's permission+ for `gh pr merge` in this environment); run CI on #420 and #421.