Skip to content

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

docs(audit): the 12 October audit

Jareerea90bffbee86
diff --git a/audit/AUDIT-2026-10-12-QUICK.md b/audit/AUDIT-2026-10-12-QUICK.md
new file mode 100644
index 00000000..c83f2adf
--- /dev/null
+++ b/audit/AUDIT-2026-10-12-QUICK.md
@@ -0,0 +1,208 @@
+# Concat audit, 12 Oct 2026 — the quick version
+
+Same audit as `AUDIT-2026-10-12.md`, one screen. Each bar has 10 slots:
+🟢 filled = good, ⚪ empty = missing. Severity: 🔴 critical · 🟠 high · 🟡 medium · ⚪ low.
+Arrows are against the 11 Oct audit. Every crate was read by one of four reviewers; the headline findings were re-read by hand.
+
+## 🎯 Verdict
+
+**`main` is green; 0.2.7 is not ready to tag.** In the 39 commits since the 11 Oct audit, its top step landed in full:
+- **CI names its own hangs.** Per-test timeouts showed which suites hung on Windows, and those suites now skip on the software GPU.
+- **Downloads and actions are pinned.** Every desktop download is pinned by digest and every action by SHA.
+- **Release bundles must start.** Every desktop bundle is launched where it was built before the release can use it.
+- **The window is fixed:** the last three dialogs are off the studio borrow, and ten tracker bugs are fixed, each with a test that fails without its fix.
+- **25 languages:** each has a font that covers its script.
+
+**But no release run has passed:**
+- **rc.1:** all six desktop builds failed.
+- **rc.2:** Linux arm64 crashed on exit. That is fixed in `c06560e`, but no release run has tried the fix yet.
+- **The phones' new launch checks have never run.**
+- **The final 0.2.7 notes would be three lines long.**
+
+Windows is green because the tests that hit its slow shader compile no longer run there. On an iPhone, Import does nothing.
+
+Count: 🟠 ×4, 🟡 ×17 new, no 🔴.
+
+## 📏 Size
+
+| | Lines | Δ vs 11 Oct |
+|---|---:|---:|
+| Rust, all crates | 90 362 | +915 (+2 007 like for like; concat-perf removed) |
+| Slint UI | 29 901 | +111 |
+| WGSL, packages | 1 764 | 0 |
+| Tests (`#[test]`) | 748 | +31 like for like |
+| Largest file `concat/src/studio.rs` | 10 866 | +177 |
+
+Checks today:
+- fmt clean, clippy 0 warnings.
+- **745 tests passed, 0 failed, 1 ignored**, on Metal.
+- 25 locales, 0 missing.
+- 23 of 23 models have their digests.
+- CI on `main` green, 6 pushes running.
+
+## 📊 The whole app
+
+| Aspect | Bar | 11 Oct | Today |
+|---|---|:-:|:-:|
+| 🧹 Code quality | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 |
+| 🏛️ Architecture | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | 8 |
+| 🧭 Design philosophy | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 |
+| 🔧 Maintainability | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 |
+| 😊 User likeability | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7 | 7 |
+| ⚡ Performance | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 |
+| 📈 Scalability | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 |
+| 🔐 Security | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5 | 6 ↑ |
+| 🧪 Testing | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 |
+| 📚 Docs | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 |
+| 🤖 CI and release | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 4 | 6 ↑ |
+
+- **Security up one:** downloads pinned by digest, actions by SHA, the release tag passed through `env:`, Windows signing by thumbprint.
+- **CI up two:** a green gate, hangs that name themselves, and bundles that must start. It is held at 6 because:
+ - two release candidates have been spent;
+ - phones are never built on a push to main;
+ - the bundle workflow can only be run by pushing a tag.
+
+## 🧱 Per crate (average of its scales)
+
+| Crate | Lines | Δ | Bar | Avg | One thing |
+|---|---:|---:|---|:-:|---|
+| ★ locales + `locales.py` | – | – | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 7.7 | 🟢 25 languages, a glyph test per locale · 🟡 a missing key does not fail CI |
+| concat-core | 2 897 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.4 | 🟢 untouched · ⚪ `MAX_SIDE`'s doc now wrong; two dead items |
+| concat-vision | 2 760 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 7.3 ↓ | 🟢 untouched · 🟡 CPU-only on shipped Windows; the Intel Mac's runtime may refuse it |
+| concat-effects | 7 256 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.9 ↓ | 🟢 123 packages pass on Metal · 🟠 none compiled on D3D12 in CI any more |
+| concat-project | 10 285 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.9 | ⚪ no commits; `AddClip` places the whole media; tests in one file |
+| concat-media | 10 066 | +212 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.8 | 🟢 nothing vendor-specific opened at launch · 🟡 an HEVC export still opens NVENC first (#309 moved) |
+| concat-render | 7 723 | +273 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.8 ↑ | 🟢 bounded waits, error scopes, phone limits · 🟡 "30 s" is 2.15 s on GLES; `combine` unscoped |
+| concat-server | 1 393 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.6 | 🟡 `stop` on `0.0.0.0` on Windows, as before |
+| concat-text | 2 607 | 0 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.6 | ⚪ no commits |
+| concat-speech | 3 307 | +198 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.5 | 🟢 a typed error and one CPU retry, tested · 🟡 Windows ships no DirectML, the crate still asks for it |
+| concat (Slint) | 29 901 | +111 | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 6.5 ↓ | 🟢 a dropdown scrolls under a finger · 🟡 three primitives still break their bindings; the captions sheet re-broke |
+| concat-export | 4 039 | +168 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.3 ↑ | 🟢 device loss tested end to end · 🟡 a validation error reads as a lost GPU |
+| concat-api | 2 352 | 0 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.1 | 🟡 captions from a clip has no verb either; the `gpu` capability can never appear |
+| concat-android | 398 | 0 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.1 | ⚪ 0 tests; the emulator check is uncommitted |
+| concat-cli | 514 | 0 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6.0 | ⚪ no logger, no signals |
+| concat-host | 10 842 | +202 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5.8 | 🟢 titles painted once; iPhone directories · 🟡 24 held stills undo it mid-drag; saves share a scratch file |
+| concat (Rust) | 23 923 | +954 | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 5.6 ↑ | 🟢 every dialog through one seam; 14 tests on pure plans · 🟠 iPhone Import does nothing |
+
+## 🎛️ Per feature (rows that moved, and the weakest)
+
+| Feature | Bar | 11 Oct | Today | State |
+|---|---|:-:|:-:|---|
+| Languages | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 7 | 8 ↑ | ✅ 25, with faces and READMEs; no right-to-left layout |
+| Export | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | 8 | ⚠️ a stuck GPU ends the export now; a Windows blur or glow can still sit for minutes, uncancellable |
+| Timeline editing | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | 8 | ✅ a click no longer moves a clip; Detach says why |
+| Titles and text | 🟢🟢🟢🟢🟢🟢🟢🟢⚪⚪ | 8 | 8 | ⚠️ painted once; past 24, they drop out of the monitor during a drag |
+| Auto-captions | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | 8 | 7 ↓ | ⚠️ ★ on the clip's menu; the sheet can transcribe a clip it is not showing |
+| ★ Command-line launch | 🟢🟢🟢🟢🟢🟢🟢⚪⚪⚪ | – | 7 | ✅ `--project`, `--version`, `--smoke`; any other argument refuses to start |
+| Phone shell | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 7 | 6 ↓ | ❌ iPhone Import does nothing; a failed start vanishes |
+| Cutout | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 7 | 6 ↓ | ⚠️ CPU-only on Windows; at risk on Intel Mac |
+| Export on a phone | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 | ⚠️ phone limits right; #286 fix uncommitted; unseen on a phone |
+| Hardware encode | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | 6 | 6 | ⚠️ #309 moved from launch to the first HEVC export |
+| ★ Bundle launch checks | 🟢🟢🟢🟢🟢🟢⚪⚪⚪⚪ | – | 6 | ⚠️ well built; no target through every check yet |
+| Enhance | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 6 | 5 ↓ | ⚠️ CPU on Windows; copies in `cache/` |
+| Remote API | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 | ⚠️ |
+| Self-update | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 | ⚠️ no signature check |
+| Windows text | 🟢🟢🟢🟢🟢⚪⚪⚪⚪⚪ | 5 | 5 | ⚠️ unseen on Windows |
+| Accessibility | 🟢🟢⚪⚪⚪⚪⚪⚪⚪⚪ | 2 | 2 | ❌ |
+
+Every other feature is unchanged; the full table is in the audit, section 4.
+
+## 🚨 Findings
+
+🟠 **High**
+1. **The final 0.2.7 notes would be three lines long.** `release.yml:170` takes `v0.2.7-rc.2` as the previous tag, and the CHANGELOG is gone. Adding `--exclude 'v*-*'` gives `v0.2.6`.
+2. **On an iPhone, Import does nothing and says nothing.** Only Android installs a file picker (`platform.rs:689-691`).
+3. **Phones are never built on a push to `main`.** The new phone checks have never run, and once committed they gate `publish`.
+4. **A Windows blur or glow export can sit for minutes compiling shaders (#297).**
+ - Nothing bounds or cancels the compile.
+ - The shaders are compiled again on every export.
+ - CI no longer runs the suites that showed it; they skip and report PASS.
+
+🟡 **Medium**
+- **Titles and saves:**
+ - Past 24 titles, the early ones drop out of the monitor during a drag, and every title is re-treated each frame (5).
+ - Two saves at once share one scratch file (6).
+- **Phones:**
+ - On GLES the "30 s" device wait is about 2.15 s, so a heavy phone frame fails the export (7).
+ - A phone that cannot start just vanishes; an iPhone's Documents folder is hidden from the Files app (12).
+- **The captions sheet** can transcribe a clip other than the one it shows (8).
+- **GPU errors and encoders:**
+ - An HEVC or AV1 export still opens NVENC first, so #309 moved from launch to export (9).
+ - Any validation error reads as "lost the GPU 3 times" (10).
+ - Transitions' `combine` is unscoped and a refused effect is silently missing from the export (11).
+- **Locales:** `locales.py --check` passes with keys missing (13).
+- **Release notes:** they carry CI-internal lines and the DXC pair, one of which reverts the other (14).
+- **ONNX Runtime:**
+ - Windows ships the CPU ONNX Runtime while speech asks for, logs and retries DirectML (15).
+ - On an Intel Mac, ONNX Runtime 1.23.2 is likely below ort's API (16).
+- **Launch checks:**
+ - `--smoke` stops before the window exists (17).
+ - The instruction-set limits are still provisional, and the emulated old CPU runs `--version` only (18).
+- **Release tooling:**
+ - The bundle workflow can only be run by pushing a tag (19).
+ - `CODE_SIGNING.md` contradicts the workflow (20).
+ - Dependabot's cargo group is red by design over sherpa (21).
+
+⚪ **Low**
+- **Phone and launcher:**
+ - The 4 px drag threshold is below a finger's slop (22).
+ - #151 comes back after Browse (23).
+- **Timeline and launch:**
+ - Detach unmutes a video the user muted on purpose (24).
+ - Any unknown argument refuses to start, so Linux "Open with" through `%U` shows an error box (25).
+- **Process and CI:**
+ - Ten "Fixes #N" commits; the push closed #299, which was reopened by hand (26).
+ - The concurrency group cancels pending runs on `main`, and 7 pushes have no CI (28).
+ - `ubuntu-latest` becomes Ubuntu 26 on 19 Oct (29).
+- **Rendering:**
+ - Right-to-left languages are laid out left to right (27).
+ - The live-title ceiling is fixed at 512 MiB rather than sized to the machine (30).
+ - The 8-bit fallback clips HDR (31).
+- **API, server and CLI:** the `gpu` capability can never appear; a gRPC stream can hold `stop`; CLI `render` writes in place (32).
+- **Docs:** three misattached doc comments, the DirectML docs, the README badge and MCP, `ARCHITECTURE.md`'s known issues (33).
+- **PR hygiene:** #304 is still open; #314 needs a re-run; 11 contributor PRs are unreviewed (34).
+
+## 🧭 What the 11 Oct audit asked, and what landed
+
+✅ **Landed:**
+- Windows hangs named, GPU suites skipped on WARP, engine job timed out, toolchain action pinned (no tag yet).
+- The three dialogs.
+- Digests, SHA pins, `env:` for the tag, signing secrets declared, the real Android key used twice.
+- The locale check prints anywhere and counts English.
+
+🟡 **Partly:** uncaptured errors are counted and scoped, with half of it left.
+
+❌ **Not landed:**
+- Relink, cache clear and trials off the UI thread.
+- The paste.
+- `Server::stop` and `catch_unwind`.
+- The pinned-frame budget and its kin.
+- The literals through `I18n`.
+- Anything seen on real Windows or a real phone.
+
+## ▶️ Next, in order
+
+1. **Release:**
+ - Fix the previous-tag lookup.
+ - Finish and commit the uncommitted phone work, on its own:
+ - the doc line;
+ - a patient final drain;
+ - `smoke-ios.sh` checks the exit status;
+ - a `trap` in `smoke-android.sh`;
+ - "Refs #286".
+ - Run `mobile.yml` by hand.
+ - Tag rc.3 and let every target finish.
+ - Tag 0.2.7, with the draft's notes edited by hand.
+2. An iPhone document picker (or a notice); a failed start shown on a phone; the two plist keys.
+3. Make `Select`, `SegmentedControl` and `Switch` controlled.
+4. Held stills stamped per pass; one scratch name per save.
+5. Windows shaders: shared pipelines, cancel between builds, a visible skip, a scheduled real-GPU job, then DXC again on real hardware.
+6. GPU errors: refused vs lost, `combine` scoped, patient GLES waits; a sentinel and a setting for vendor encoders.
+7. CI:
+ - `push` on `mobile.yml`, `workflow_dispatch` on `build-app.yml`;
+ - `--smoke` through real frames, instruction-set limits from rc.2's counts;
+ - pin `ubuntu-24.04` before 19 Oct;
+ - Dependabot's sherpa exclusion;
+ - `locales.py` fails on missing keys.
+8. Honest ONNX providers on Windows and Intel Mac; `CODE_SIGNING.md` in line with the workflow.
+9. Then the 11 Oct steps still open: the paste, the server, the pinned-frame budget, and the 28 Sept list.
diff --git a/audit/AUDIT-2026-10-12.md b/audit/AUDIT-2026-10-12.md
new file mode 100644
index 00000000..cdef6f8a
--- /dev/null
+++ b/audit/AUDIT-2026-10-12.md
@@ -0,0 +1,863 @@
+# Concat audit, 12 October 2026
+
+Where the codebase stands one day after the 11 October audit, at
+`c06560e` (12 October), 39 commits past the last audit's `7223d64` and
+110 past the 0.2.6 tag, with two release candidates (`v0.2.7-rc.1`,
+`v0.2.7-rc.2`) tagged and failed in between. Same method, same scales,
+every finding from the 11th (and what it carried from the 9th, 4th and
+28th) re-verified against today's tree. Every finding names a file and a
+line. The sibling `AUDIT-2026-10-12-QUICK.md` is this audit as a
+one-screen scorecard.
+
+**How it was done.** Four reviewers each read one part of the tree: the
+window's Rust, Slint and locales; host, server, API and CLI; media,
+render, export and core; project, effects, text, vision, speech, Android,
+CI, release and docs. Each re-verified the 11 October findings in its part
+against today's lines, read the commits since the audit as commits (does
+each do what its message says, with what test, missing what), scored its
+part and reported new findings with locations. The headline items (the
+release notes' previous tag, the iPhone's Import, the pool's 24 held
+stills, the shared save scratch, the hard-coded hardware encode, the
+locale check's missing keys, the auto-closed issues) were re-read by hand,
+and where a command could show it, run. Line, test and `unsafe` tallies
+use the 11th's counting (Rust lines are `src/` only, without `build.rs`
+and `tests/` directories; the bracket is the count outside test modules),
+which the reviewers reproduced exactly on the 11th's tree before counting
+today's. `cargo fmt --check`, `cargo clippy --workspace --all-targets`,
+`cargo test --workspace --no-fail-fast`, `scripts/locales.py --check` and
+`scripts/models.py --check` were run on this machine (an Apple silicon
+Mac, macOS 26, Metal) today. The app itself was not launched; no phone or
+release bundle was built here.
+
+**The working tree is not clean.** Uncommitted on this machine and
+intended for 0.2.7: a send-and-receive retry for encoders that answer
+"try again" (`concat-media/src/encode.rs`, aimed at Android's MediaCodec,
+#286), Android-emulator and iOS-simulator launch checks
+(`scripts/ci/smoke-android.sh`, `smoke-ios.sh`, +197 lines of
+`mobile.yml`), and an error exit when the event loop fails
+(`concat/src/lib.rs`). This audit scores `HEAD`; section 6 reviews the
+uncommitted work separately, because it gates the release once
+committed. The test run below includes it.
+
+**What the 39 commits are.** All on 11 and 12 October; 33 by the
+maintainer, 6 by three contributors (Bengali, Vietnamese, `--project` on
+launch). In groups:
+
+- *The release pipeline* (d118484, b161c1d, 2d3a43e, c06560e): every
+ bundle launched before upload (`--version`, `--smoke`, an emulated Ivy
+ Bridge, an instruction scan), every download pinned by digest, every
+ action by SHA, nextest with per-test timeouts, draft releases, Windows
+ signing by thumbprint; then the fixes the two release candidates found.
+- *Windows* (be1b882, 3178b78, 0a61b65): DXC in, then out the same day;
+ the GPU suites that took minutes on WARP skip there; no Microsoft Store;
+ the installer in the app's languages. `main` is green because of these.
+- *Bugs from the tracker* (132f68d, 34fe37a, 590e483, f1ccdfe, 049823f,
+ e5d8197, e1b64a4, b4adc57, 50c18ee): iPhone projects, settings and logs
+ (#299); a second Untitled project (#151); a click that moved a clip
+ (#247); Detach that said nothing (#310); no sound band on a silent file
+ (#283); the last three dialogs off the studio borrow; the inspector's
+ section strip (#250) and a title's words field (#139); titles painted
+ once (#261); Chatterbox read again on the CPU (#291); an export the GPU
+ stops answering ends, and a phone's GPU draws within its limits (#297,
+ #312); no hardware encoder opened at launch (#309); a dropdown scrolls
+ under a finger.
+- *Features* (2d3e651, 5982dc7, 45c061f): Generate captions on a clip's
+ menu and a "Use voice from timeline" switch (#302); `--project` on launch.
+- *Languages* (30a1742, ad59b61, c65295b, 8a628c6, c0c8827, fa51932,
+ 5c60bfd, 9737836): 14 languages to 25, a README in each, five new Noto
+ fallback faces with a glyph test per locale, a locale check that prints
+ on Windows and counts English.
+- *Housekeeping* (1d22c12, 7c5cb3d, 0714fb8, 4affa32, 2d27766, af62333,
+ 91a6df0, the 11 October audit itself): the CHANGELOG, concat-perf and the
+ web build removed; AGPL-only licence text; a code-signing policy; the
+ version bumped to 0.2.7.
+
+**Scale.** 0 to 10. 8 and up means "keep doing this"; 5 to 7 means
+"works, with known debt"; 4 and under means "will bite at the next order
+of magnitude, or already bites".
+
+---
+
+## 1. The numbers
+
+### 1.1 Size, tests and hygiene per crate
+
+Deltas are against the 11 October table. concat-perf (1 092 lines, 2
+tests) was removed in 7c5cb3d; like for like the tree grew by 2 007 lines.
+The 11th's concat-render row counted `src/gpu/tests.rs`, a `#[cfg(test)]`
+module (`gpu.rs:3560-3561`), as non-test code: its "(68)" unwraps and "3"
+panics were 59 and 3 in tests. Corrected below; the totals' brackets
+follow.
+
+| Crate | Rust lines | Δ | Slint / WGSL / Java | Tests (files) | `unsafe` | `unwrap`/`expect` (non-test) | `panic!` family (non-test) |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| concat (window) | 23 923 | +954 | 29 901 slint (+111) | 87 (20) | 13 | 37 (4) | 0 |
+| concat-host | 10 842 | +202 | | 89 (19) | 4 | 249 (2) | 0 |
+| concat-project | 10 285 | 0 | | 114 (1) | 0 | 356 (11) | 6 |
+| concat-media | 10 066 | +212 | | 100 (11) | 38 | 204 (23) | 0 |
+| concat-render | 7 723 | +273 | shaders inline in `gpu.rs` | 91 (5) | 1 | 86 (9) | 0 |
+| concat-effects | 7 256 | 0 | 1 764 wgsl | 72 (9) | 0 | 134 (5) | 2 |
+| concat-export | 4 039 | +168 | | 34 (3) | 0 | 31 (0) | 0 |
+| concat-speech | 3 307 | +198 | | 21 (3) | 3 | 23 (2) | 0 |
+| concat-core | 2 897 | 0 | | 49 (7) | 0 | 26 (4) | 2 |
+| concat-vision | 2 760 | 0 | | 31 (8) | 0 | 22 (2) | 0 |
+| concat-text | 2 607 | 0 | | 24 (1) | 1 | 67 (1) | 0 |
+| concat-api | 2 352 | 0 | | 21 (2) | 0 | 23 (0) | 0 |
+| concat-server | 1 393 | 0 | | 12 (4) | 0 | 28 (0) | 0 |
+| concat-cli | 514 | 0 | | 3 (1) | 0 | 2 (2) | 0 |
+| concat-android | 398 | 0 | 291 java | 0 (0) | 10 | 1 (1) | 0 |
+| ~~concat-perf~~ | – | −1 092 | | – | – | – | – |
+| **Total** | **90 362** | **+915** | **~31 960** | **748 (96)** | **70** | **1 289 (66)** | **10** |
+
+The window took most of it again (+954: `launch.rs`, the untitled naming,
+the detach plan, the captions sources, the fonts table, `leave()`), then
+host's title cache (+202), render's bounded waits and phone limits (+273),
+media's encoder seams (+212) and speech's CPU retry (+198). `studio.rs` is
+10 866 lines (+177; 45 % of the window's Rust): `impl Studio` from `:1962`
+to `:10419` is 8 458 lines and about 220 methods over the same 93 fields,
+and `run()`'s body moved into `fn start` (`lib.rs:161-1932`), 1 772 lines.
+One `unsafe` was added (`leave`'s `libc::_exit`, with a SAFETY line); all
+70 carry one, and the one that is not true as written (`platform.rs:
+347-350`) is as it was. 31 tests landed on 2 007 lines, every one on a
+pure helper that fails without its fix.
+
+Other facts: 123 effect packages, unchanged; 25 locale files (en and 24)
+at 1 020 keys each, 0 missing and 0 stale, with every `{n}` placeholder
+matching English (checked by hand; the script does not); 993 values still
+the English word for word, as the script now counts it: fr 150, de 142,
+it 125, pt-BR 125, es 123, tr 109, ko 99, nl 44, hr 31, id 28, pl 10,
+vi 4, fa 2, bn 1, and ten locales at 0. 39 commits since the last audit
+(161 files, +21 465 / −3 445, most of it locales and READMEs); 32 tags,
+latest `v0.2.7-rc.2`; the last release is `v0.2.6`, 110 commits behind
+`main`. 82 open issues (89 on the 11th) and 19 open pull requests, 11 of
+them from one contributor, unreviewed since 6 October.
+
+### 1.2 The checks, run today
+
+| Check | Result |
+|---|---|
+| `cargo fmt --all -- --check` | clean |
+| `cargo clippy --workspace --all-targets` | 0 warnings |
+| `cargo test --workspace --no-fail-fast` | **745 passed, 0 failed, 1 ignored**, on Metal: the GPU suites ran for real |
+| `scripts/locales.py --check` | 25 locales, 1 020 keys, 0 missing, 0 stale; warns 993 English values; **would pass with keys missing** (finding 13) |
+| `scripts/models.py --check` | 23 models, 3.37 GB, mirrored on `models-v1` |
+| `cargo audit` | not installed on this machine; not run |
+| CI on `main` | **green** at `2d3a43e` and on every completed push since `3178b78` (six in a row, the first since 4 October); `c06560e` in progress at the time of writing |
+| Release, `v0.2.7-rc.1` | CI passed; **all six desktop builds failed** (Windows smoke silent for 15 minutes; Linux and macOS smoke scripts without the exec bit; Intel Mac build past 60 minutes; Windows on ARM's LLVM installer); Android and iOS passed |
+| Release, `v0.2.7-rc.2` | Linux arm64's `--smoke` logged `smoke: ok` and **segfaulted on its way out** (139); the rest cancelled by hand at 20:07; Android passed and was signed with the real key |
+
+**How `main` went green.** d118484 moved CI to nextest
+(`src/.config/nextest.toml:12-40`: slow at 60 s, killed at 5 minutes,
+JUnit kept on failure), which is what the 11th asked for, and the next
+Windows run named its hang at once: `every_package_probe_holds`,
+`every_shader_package_renders_at_its_defaults`,
+`the_gaussian_blur_is_a_direct_gaussian` and `every_edit_still_exports`,
+each past its limit on WARP (run 38209190014). That is the FXC compile of
+the blur and glow loops, the same cost #297 is about. be1b882 switched
+D3D12 to DXC to cure it; two GPU tests failed outright under DXC on WARP,
+and 3178b78 reverted it 38 minutes later and made those suites skip on a
+software adapter (`concat-render/src/gpu/tests.rs:41-48`,
+`concat-host/tests/export.rs:1550-1569`). Windows has run 740 tests green
+in 15 to 17 minutes since. The cost is that nothing in CI compiles an
+effect package through D3D12 any more (finding 4).
+
+**What has passed in a release run.** No desktop target has run every
+check to the end. Linux x86_64 came closest (host smoke, the emulated Ivy
+Bridge and the instruction scan, then cancelled in the clean-distribution
+installs). The Windows smoke has never got past its first line; the macOS
+smoke has never run; the Intel Mac has never finished a build. 2d3a43e
+(Windows smoke and emulator split, 100 minutes for the Intel Mac) and
+c06560e (`_exit` on Linux) are unproven until rc.3.
+
+### 1.3 Tests
+
+`cargo test --workspace --no-fail-fast` ran to completion here: 745
+passed, 0 failed, 1 ignored (the one ignored is concat-speech's model
+test). It ran on the working tree, so it includes the uncommitted
+encoder retry and its one test. The attribute count (748) is higher than
+the run because a few tests are compiled for other platforms only. On
+this Mac the GPU suites that skip on Windows' WARP ran on Metal and
+passed, including all 123 packages at their defaults, which is the only
+place today they were run on a real GPU. There is no performance table:
+concat-perf is gone (7c5cb3d), and nothing replaces it.
+
+---
+
+## 2. Scorecard for the whole app
+
+| Aspect | 11 Oct | Today | One line |
+|---|:-:|:-:|---|
+| Code quality | 6 | 6 | fmt green, clippy clean, four `expect`s in 24k lines of window code, every new helper pure and tested; against that, three doc comments stacked or misattached, ten commits that say "Fixes" against the repository's rule, and d118484 claiming a Windows suite that could finish when it could not. |
+| Design architecture | 8 | 8 | Held. Every native dialog now goes through one seam; platform rules (`dirs::convention`, `windows_kind`, `location_on`, `encodable_among`) are pure functions of their inputs, so each platform's rule is tested on every runner. |
+| Design philosophy | 6 | 6 | The week's fixes go to the cause where they can (bounded waits, an honest revert, a launch check per bundle); but two binding-break fixes treat the symptom while the three primitives that cause it stay, and the captions sheet brought the bug back a day later. |
+| Maintainability | 5 | 5 | `studio.rs` 10 866 lines (+177), `start()` 1 772, `gpu.rs` 3 561, `concat-export/src/lib.rs` 3 210, `encode.rs` 2 173; concat-project's 114 tests in one file. |
+| User likeability | 7 | 7 | 25 languages with a face for every script; captions from the clip's menu; a click no longer moves a clip; Detach says why; the iPhone can make a project. Against that: on an iPhone Import does nothing; the captions sheet can transcribe a clip it is not showing; titles drop out of the monitor during a drag; a Windows blur or glow export can sit for minutes. |
+| Performance | 6 | 6 | #261's title cache is right, but the pool's 24 held stills undo it mid-gesture and re-treat every title each frame; the FXC compile is back on Windows and repeats on every export; concat-perf is gone. |
+| Scalability | 5 | 5 | Unchanged: the fd per connection, the one Hub thread, the per-sample mixer scan, the per-edit bin clone. |
+| Security | 5 | 6 | Every desktop download pinned by digest (`scripts/ci/deps.lock`, `fetch.sh`), every action by SHA, the tag validated and passed through `env:`, Windows signing by thumbprint, the Android key proven twice. Against that: mobile downloads still undigested, the signing policy and the workflow disagree, the update chain as it was. |
+| Testing | 6 | 6 | 31 tests, each failing without its fix; per-test timeouts; every desktop bundle launched before upload. Against that: the skipped GPU suites report PASS, `--smoke` stops before the window exists, the emulated old CPU runs `--version` only, concat-android has no tests, and there is no performance gate. |
+| Documentation | 5 | 5 | The CHANGELOG's staleness ended by deleting it, and the notes that replace it would give 0.2.7 three lines; `CODE_SIGNING.md` contradicts the workflow; `ARCHITECTURE.md`'s known issues were stale within a day; the README still sells an MCP server. |
+| CI and release | 4 | 6 | `main` green six pushes running; a hang names itself in minutes; draft releases that need every bundle launched. Against that: two release candidates spent and no desktop target through every check, phones never built on a push, the bundle workflow runnable only by a tag, and the final notes' previous tag wrong. |
+
+---
+
+## 3. Scores per crate
+
+Columns: code quality, maintainability, scalability, performance,
+architecture, design philosophy, security, testing. A dash means the
+scale does not apply. Arrows are against 11 October.
+
+| Crate | CQ | Maint | Scale | Perf | Arch | Phil | Sec | Test | Note |
+|---|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|---|
+| concat-core | 8 | 8 | 6 | 7 | 8 | 8 | 7 | 7 | No commits. `MAX_SIDE`'s doc (`frame.rs:28-31`, "the downlevel limit every GPU meets") is now contradicted by render's own `max_side`; dead `Project` and `source_duration`; `Neg` and `ceil` on `i64::MIN`. |
+| concat-project | 8 | 6 | 6 | 6 | 8 | 7 | 6 | 8 | No commits. `AddClip` places the whole media (`commands/clips.rs:34-38`), the speed curve grows over a neighbour (`properties.rs:285-308`), `DOCUMENT_VERSION` is 1 (`doc.rs:41`), all tests in one 4 294-line file. |
+| concat-media | 7 | 6 | 5 | 7 | 8 | 7 | 7 | 7 | b4adc57's seams are pure with injected `opens` and environment, and its tests panic if anything is opened; but an HEVC or AV1 export still opens NVENC first (finding 9). The uncommitted retry is right by FFmpeg's contract. |
+| concat-render | 7 | 5 | 6 | 7 | 8 | 7 | 7 ↑ | 7 | Bounded waits, per-frame error scopes, the adapter's own limits, an 8-bit fallback tested against the half-float picture. But on GLES "30 seconds" is about 2.15 (finding 7); `combine` and refused passes are unscoped (finding 11); four docs still promise a CPU fallback. |
+| concat-effects | 8 | 7 | 6 | 5 | 8 | 8 | 6 | 7 ↓ | No commits. The package suites no longer run on Windows CI: the 123 shaders are proven on lavapipe, Metal and nothing on D3D12. |
+| concat-export | 7 | 4 | 6 | 5 | 7 | 7 | 7 | 7 ↑ | The device-loss recovery is tested end to end through `render_on`; progress at least once a second; a phone-sized ceiling with a reason. A deterministic validation error now reads as a lost device (finding 10); `lib.rs` 3 210 lines. |
+| concat-host | 7 | 6 | 5 | 6 | 6 | 6 | 4 | 6 | The #261 LRU and the iPhone directories are small, pure where they can be, and tested; the pool's FIFO undoes the first mid-gesture (finding 5) and overlapping saves share a scratch file (finding 6); the update chain as it was. |
+| concat (Rust) | 6 | 4 | 5 | 6 | 7 ↑ | 6 | 6 | 5 ↑ | Every dialog through one seam; 14 tests on pure plans (`detach_plan`, `moved_any`, `location_on`, `launch::parse`); `leave()`. Against: the iPhone's Import (finding 2), three misattached docs, the paste unchanged, `studio.rs` +177. |
+| concat (Slint) | 7 | 7 | 6 | 7 | 6 ↓ | 6 | – | – | Two binding-break fixes and a finger-scrollable Select; but `select.slint:127`, `segmented-control.slint:43` and `controls.slint:559` still assign their own value, about 33 one-way bindings remain, and the captions sheet re-introduced the bug (finding 8). |
+| locales and `locales.py` ★ | 8 | 8 | 7 | – | 8 | 8 | – | 7 | 25 languages with READMEs in parity, a face for every script with a charmap test per locale, English counted per locale. Missing keys do not fail; no placeholder check; no right-to-left layout. |
+| concat-api | 8 | 7 | 5 | 5 | 8 | 4 | 5 | 7 | No commits. The parity promise (`message.rs:23-26`) is behind again: captions' two new entry points are window-only. `export.run` overwrites; the `gpu` capability can never appear. |
+| concat-server | 7 | 8 | 5 | 6 | 8 | 8 | 5 | 6 | No commits. The wildcard-stop hang, the fd per connection, seats before auth; a gRPC `Events` stream can hold `stop`. |
+| concat-cli | 7 | 8 | – | – | 7 | 6 | 6 | 2 | No commits. No logger; `serve` parks with no signal handling; `render` writes in place. |
+| concat-speech | 7 | 5 | 6 | 6 ↓ | 7 ↑ | 7 | 8 | 6 | A typed `ReadError` and a once-only CPU retry, with two pure tests; but Windows ships the CPU ONNX Runtime while the crate asks for, logs and retries DirectML (finding 15). |
+| concat-vision | 8 | 8 | 5 | 5 ↓ | 9 | 8 | 8 | 7 | Code unchanged; cutouts, the brush and enhance are CPU-only on shipped Windows, and the Intel Mac's runtime may not serve ort's API at all (finding 16). |
+| concat-text | 7 | 6 | 5 | 6 | 8 | 8 | 7 | 6 | No commits. |
+| concat-android | 7 | 7 | – | 7 | 7 | 7 | 7 | 1 | No code commits; three bridge crash paths; 0 tests. The uncommitted emulator check would be its first. |
+
+---
+
+## 4. Scores per feature
+
+Status: **done**, **partial** (works with a known gap), **thin** (present,
+weak). Rows that changed since the 11th carry a reason; ★ marks a row new
+this audit.
+
+| Feature | Where | Status | 11 Oct | Today | Why |
+|---|---|---|:-:|:-:|---|
+| Launcher, new project, recents | `start.slint`, `panes/start.rs` | done | 8 | 8 | Open off the borrow; "Untitled project N" (#151) and the iPhone's location (#299), with four tests. A Browse brings #151 back (finding 23); the Files app cannot see the iPhone's folder (finding 12). |
+| Import, bin, drag and drop | `media-pane.slint`, `host.rs:377-400` | done | 8 | 8 | Menu and bin Import through `pick_files_then`. On an iPhone it does nothing (finding 2). |
+| Timeline editing | `lanes.slint`, `studio.rs` | done | 8 | 8 | A click no longer moves a clip (#247, tested), Detach says why (#310, tested), no band on a silent file (#283). The 4 px threshold is under a finger's slop (finding 22); Detach unmutes a video muted on purpose (finding 24). |
+| Copy and paste | `studio.rs:6855-6966` | partial | 6 | 6 | Unchanged. |
+| Volume line | `lanes.slint` | done | 7 | 7 | Absent on a file with no sound, consistently. |
+| Undo and redo | `editor.rs` | done | 8 | 8 | Unchanged. |
+| Transitions | `plan.rs`, `gpu.rs:3450-3518` | partial | 7 | 7 | `combine` runs outside the new error scopes. |
+| Effects and filters | `concat-effects/packages/` | done | 8 | 8 | Code unchanged; on Windows the blur and glow compile cost of #297 ships again (finding 4). |
+| Colour and HDR | `model.rs`, `gpu.rs`, `encode.rs` | done | 7 | 7 | The 8-bit fallback on a phone clips HDR with a log line only. |
+| Colour grading | `concat.color-wheels` | done | 7 | 7 | Unchanged. |
+| Titles and text | `concat-text`, `titles.rs` | done | 8 | 8 | #261: a timeline past the budget is painted once (049823f, tested); more than 24 titles drop out of the monitor during a drag (finding 5). |
+| Shapes / Stickers | `media-pane.slint`, `clips.rs` | partial | 6 | 6 | Unchanged. |
+| Keyframes, animations | `model.rs` | done | 7 | 7 | Unchanged. |
+| Crop, transform, blend, masks | `plan.rs` | done | 7 | 7 | Unchanged. |
+| Cutout, smart brush | `segment.rs` | done | 7 | 6 ↓ | CPU-only on shipped Windows; the Intel Mac's ONNX Runtime may refuse its first session (finding 16). |
+| Enhance | `enhance.rs` | thin | 6 | 5 ↓ | The same runtime as cutout; copies still in `cache/`. |
+| Auto-captions | `captions.slint`, `panes/captions.rs` | done | 8 | 7 ↓ | ★ On a clip's menu, with a source list and a switch (#302); but the sheet can show one clip and transcribe another, and show one placement and use another (finding 8). |
+| Text-to-speech | `tts.rs`, `chatterbox.rs` | done | 7 | 7 | The #291 retry helps CoreML; on Windows DirectML is not shipped at all, so the retry doubles a real CPU failure (finding 15). |
+| Voiceover | `record.rs` | partial | 6 | 6 | A crashed take is still lost. |
+| VU meters | `meters.rs` | partial | 6 | 6 | No commits. |
+| Templates | `templates.rs` | done | 7 | 7 | Unchanged. |
+| LUT import | `cube.rs` | done | 8 | 8 | Off the borrow now. |
+| Custom and community packages | `catalogue.rs` | done | 7 | 7 | Unchanged. |
+| Proxies | `proxy.rs` | partial | 5 | 5 | Unchanged. |
+| Export | `panes/export.rs`, `concat-export` | done | 8 | 8 | A GPU that stops answering ends or recovers the export instead of freezing it, and a slow frame moves the bar. Against that: an FXC compile on Windows that nothing bounds or cancels, a refused effect silently absent, a validation error reported as a lost device. |
+| Export on a phone | `platform.rs`, `ConcatFiles.java` | partial | 6 | 6 | GLES devices open within their own limits and draw in 8 bits where half floats fail; the uncommitted retry targets #286; a heavy frame fails after about 2.15 s (finding 7). Still not seen on a phone. |
+| Hardware encode | `encode.rs` | done | 6 | 6 | Nothing vendor-specific opened at launch (#309); the crash moves to the first HEVC or AV1 export (finding 9); "· hardware" is blank until an export has run. |
+| Playback and audio | `playback.rs` | done | 7 | 7 | Unchanged. |
+| Monitor stills | `pool.rs` | done | 7 | 7 | A dead headless compositor is replaced. |
+| Hardware decode | `hardware.rs` | done | 7 | 7 | Unchanged. |
+| Remote API | `concat-api`, `concat-server` | partial | 5 | 5 | Unchanged; one more window operation (captions from a clip) without a verb. |
+| Self-update | `updates.rs` | done | 5 | 5 | The Store branch gone; the chain as it was, and no Authenticode check although every PE is now signed. |
+| Logs | `logs.rs` | done | 7 | 7 | On an iPhone the log is now written at all. |
+| Keyboard menus, platform keys | `app.slint`, `platform.rs` | partial | 6 | 6 | Unchanged. |
+| Save Video Frame As, Save Audio As | `studio.rs` | done | 8 | 8 | Unchanged. |
+| Phone shell | `ui/phone/`, `concat-android` | partial | 7 | 6 ↓ | Dropdowns scroll under a finger; captions on the clip menu. On iOS, Import is dead and a failed start is silent (findings 2, 12); on Android a failed start vanishes too. |
+| Relink missing media | `relink.rs` | done | 6 | 6 | Unchanged. |
+| Themes and accent colours | `ui/theme/` | done | 8 | 8 | Unchanged. |
+| Languages | `locales/`, `fonts.rs` | done ↑ | 7 | 8 | ★ 25 languages and 25 READMEs; five subset Noto faces with digests and a charmap test per locale; English counted per locale. No right-to-left layout for ar, he, ur, fa; seven older locales 99 to 150 English; a missing key does not fail CI. |
+| Windows text | `src/patches/i-slint-renderer-skia` | partial | 5 | 5 | Untouched; unseen on a Windows screen. |
+| Accessibility | the whole UI | thin | 2 | 2 | 2 `accessible-*` properties. |
+| Windows data folders | `dirs.rs` | done | 6 | 6 | Unchanged. |
+| ★ Command-line launch | `launch.rs`, `main.rs` | done | – | 7 | `--project`, `--version`, `--smoke`, parsed and tested; any other argument now refuses to start (finding 25); `--version` prints nothing in a Windows console. |
+| ★ Bundle launch checks | `scripts/ci/smoke-*`, `isa-scan.sh` | partial | – | 6 | Well built: an import audit on Windows, clean-distribution installs, the CUDA strace, an instruction scan. No target through every check yet; `--smoke` stops before the window; the emulated old CPU runs `--version` only; the instruction limits are provisional. |
+
+---
+
+## 5. The 11 October findings, re-verified
+
+Numbers are the 11 October audit's (its section 6).
+
+| # | Finding | Today | Evidence |
+|---|---|---|---|
+| 1 | `main` red; the Windows hang undiagnosed | **Fixed, by skipping** | nextest named four suites past their limits on WARP (run 38209190014); `heavy_gpu()` (`gpu/tests.rs:41-48`) and the glow skip (`concat-host/tests/export.rs:1550-1569`) return early on a software adapter; six green pushes since `3178b78`. What the skip hides is finding 4. |
+| 2 | Three dialogs opened under the studio borrow | **Fixed** (132f68d) | Start › Open `panes/start.rs:283-292` (answered as `StartMsg::OpenChosen`), Import LUT `studio.rs:7566-7572`, menu Import `lib.rs:1672`, bin Import and Import Font `lib.rs:710, 733`, all through `host::after_dialog` or `pick_files_then` (`host.rs:352-400`). No dialog call remains outside one. |
+| 3 | A paste does not keep its shape; a title's copy keeps a quarter | Unchanged | `place_copy` `studio.rs:6855-6966`: `AddClip` lands the whole media (`:6888-6896`, `commands/clips.rs:34-38`) then trims (`:6899-6918`); titles return early (`:6862-6887`); the clipboard is not cleared in `close_project` (`:7218`). |
+| 4 | The toolchain action on a moving branch | **Fixed** | Pinned to a SHA in every job (`ci.yml:56, 252`; `build-app.yml:201`; `mobile.yml:111, 333`). Dependabot cannot move a branch pin, so it stays frozen until moved by hand. |
+| 5 | The CHANGELOG has nothing after 0.2.6 | **Obsoleted** (1d22c12 deleted it) | A release's notes are now its commits alone, and the previous tag is found wrong (finding 1). |
+| 6 | `Server::stop` can hang on a wildcard bind on Windows | Unchanged (reasoned) | `concat-server/src/lib.rs:269-272, 281-283`; called on the UI thread (`settings.rs:557-558`). |
+| 7 | A refused thread kills the accept loop silently | Unchanged | `json.rs:128, 183`. |
+| 8 | An uncaptured GPU error is a silently wrong export frame | **Partly fixed** (e1b64a4) | `render()` now draws under OutOfMemory and Validation scopes and compares an uncaught-error counter (`gpu.rs:3393-3441`); a failed frame is redrawn on a fresh device (`concat-export/src/lib.rs:1525-1541`). The rest is finding 11. |
+| 9 | `studio.rs` keeps growing | **Worse**, slowly | 10 866 lines (+177); `detach_audio`, `moved_any` and `detach_plan` were added to it (`:10266`, `:10439`, `:10455`). The week's other new code went into panes and pure functions. |
+| 10 | The API writes where the window asks | Unchanged | `concat-api/src/lib.rs:519-520` (no `exists`), `:380-397` (no `is_sane`); a `rate_den` of 0 leaves an empty project folder (`projects.rs:234-236`). |
+| 11 | Mux scratch names collide across exporters | Unchanged, **one correction** | `concat-export/src/lib.rs:976-978`. The window's embedded API shares the window's export slot (`settings.rs:582-583`), so the collision is between processes and same-stem outputs, not between the window and its API. |
+| 12 | The export's two decode paths disagree past the end | Unchanged | `lib.rs:1484-1487` against `pool.rs:427-432, 999-1008`. |
+| 13 | `pick_rate` and variable-rate files | Unchanged | `probe.rs:174-178, 212-232`; no VFR fixture. |
+| 14 | Five things the ceilings left | **Partly fixed** | Both devices now ask for the adapter's limits (`gpu.rs:548-552`; `concat/src/gpu.rs:94-98`), a device under 8 192 is refused at export start (`gpu.rs:849-851`, `lib.rs:1321-1328`), and the window's comment is right. `frame.rs:28-31` is still wrong; `transition_in.duration` has a floor only (`model.rs:1510-1512`); `VideoSettings::or` falls back silently. |
+| 15 | The server's quieter edges | Unchanged | `grpc.rs:74-76`, `hub.rs:54-66, 149-151`, `concat-cli/src/main.rs:172-173`, `concat-api/src/lib.rs:868-871`. |
+| 16 | The cards skip and the engine job | **Partly fixed** | The engine job has `timeout-minutes: 130` (`ci.yml:41`); the no-adapter branch still skips under `CONCAT_REQUIRE_GPU` (`cards.rs:598-601`). |
+| 17 | The signing step's unverified edges | **Fixed** | The alias from `keytool -list`, `--key-pass env:`, stdin from `/dev/null` (`mobile.yml:205-220`); the Windows secrets declared (`build-app.yml:68-73`); the real key used in both release candidates. |
+| 18 | The window's smalls | Unchanged | `host.rs:357-367` (no `refresh_art`), `dock.rs:19-20`, `meters-pane.slint:198, 206, 214`, the tray's keyboard path. |
+| 19 | concat-project's tests in one file | Unchanged | `lib.rs`, 4 294 lines, all 114 tests. |
+| 20 | `locales.py` crashes on a Windows console | **Fixed** (9737836) | `scripts/locales.py:43-47`. |
+| 21 | Doc drift | **Partly fixed** | `src/README.md` dropped perf and wasm but still omits concat-text and concat-vision (`:9-21`); the CPU-fallback docs remain (`concat-export/src/lib.rs:18-20`; `gpu.rs:676-678, 1295-1296, 1644-1648`; `concat-host/Cargo.toml:37-38`; `concat/Cargo.toml:124`; `platform.rs:335, 520-523`); `tts.rs:7`; MCP at `concat-api/src/lib.rs:7` and `README.md:58, 75, 106`. |
+| 22 | Dead code in core | Unchanged | `timeline.rs:314-321, 497-514`. |
+
+Of the 22: 5 fixed, 1 obsoleted, 4 partly, 11 unchanged, 1 worse. Of what
+the 11th carried from the 9th: #6 (dialogs) is **fixed**; #21 (downloads
+and action tags) is **mostly fixed**: every desktop and CI download goes
+through `scripts/ci/fetch.sh` against `deps.lock` (24 files, FFmpeg a
+dated build rather than `latest`), every action is pinned by SHA, and
+`inputs.tag` is validated and passed through `env:` (`release.yml:50-57,
+65-83`), leaving only the phone downloads (`ffmpeg-mobile.sh:38-39`,
+`sherpa-mobile.sh:38-39, 56-57`); #24 (tests that prove the wrong thing)
+is **partly** better (device loss now driven end to end, the hardware
+decode test honest) and partly worse (two of e1b64a4's tests restate the
+code: `gpu/tests.rs:1773-1781, 1850-1856`); and the 11th's correction to
+#20 needs one of its own: `finish` caps the RIFF size only past
+`u32::MAX`, so a take within 36 bytes below it still overflows at
+`record.rs:287` (a 12-hour take; trivial). The rest (#10, #11, #12, #13,
+#14 the walk, #15, #16, #17, #19, #20, #22, #23, #25 to #29, and the 4th's
+#4, #5, #7, #17 to #19, #23, #26, S1 to S4, S6, S8, S9 partly, S14, S16,
+S17) are unchanged, at the lines the 11th gave or a few lines on.
+
+---
+
+## 6. New findings, by severity
+
+Numbers are this audit's. Each names where to look.
+
+### High
+
+**1. The final 0.2.7 release would ship three lines of notes.** The
+release finds the previous tag with `git describe --tags --abbrev=0
+--match 'v*' "$tag^"` (`.github/workflows/release.yml:170`), and the two
+release candidates match `v*`. Run here, it answers `v0.2.7-rc.2`, and
+`release-notes.sh v0.2.7-rc.2 HEAD 0.2.7` prints three lines (the `_exit`
+fix, the Windows smoke split, the captions switch); from `v0.2.6` it
+prints 82. With `CHANGELOG.md` gone (1d22c12) nothing else carries
+0.2.6 to 0.2.7, and the draft (`:179`) catches it only if someone reads
+it. Add `--exclude 'v*-*'` when the tag has no hyphen (run here, that
+answers `v0.2.6`). (verified by running)
+
+**2. On an iPhone, Import does nothing and says nothing.** The phone's
+Import (`phone/editor.slint:850-854`) and Import Font (`:1103`) reach
+`platform::pick_files_async`, which on a phone asks `FILE_PICKER`
+(`platform.rs:686-692`); only `concat-android/src/lib.rs:223` ever
+installs one, so on iOS the answer is `log::warn!("no file picker on this
+platform yet")` and nothing on screen. 34fe37a and 132f68d made an iPhone
+able to save settings and create a project; without import it cannot
+edit anything. Export is the same shape: no publisher on iOS
+(`platform.rs:629-635`), so a finished file stays in the sandbox (and
+finding 12). Install a `UIDocumentPickerViewController` picker; until
+then, `notify` instead of logging. (verified)
+
+**3. Phones are never built on a push to `main`, and the new phone checks
+have never run.** `mobile.yml`'s triggers are `workflow_call`,
+`pull_request` and `workflow_dispatch` (`HEAD:.github/workflows/
+mobile.yml:4, 26-27`); all work lands on `main` directly, so an Android
+or iOS break shows only on a contributor's PR, a hand-started run or a
+tag. The uncommitted emulator and simulator checks inherit this, and once
+committed they gate `publish` (`release.yml`'s `publish` needs `phones`),
+so their first run would be rc.3. Run `mobile.yml` by hand on `main`
+before tagging; add `push: branches: [main]`. (verified)
+
+**4. On Windows, a blur or glow export can sit for minutes compiling
+shaders, and no test watches it.** With DXC reverted (3178b78), D3D12
+compiles through FXC, which be1b882's own message says takes minutes over
+the blur and glow loops (#297, open). Pipeline creation is CPU work
+(`gpu.rs:2497-2558`), so `DEVICE_WAIT` does not bound it; cancel is read
+at the top of each frame (`concat-export/src/lib.rs:1381`) and progress
+after one (`:1551-1554`), so Cancel does nothing until the compile ends.
+It repeats on every export: the window lends `monitor.sibling()`
+(`panes/export.rs:516`), whose `drawing_in` starts an empty `shaders` map
+(`gpu.rs:836-846, 1291`), so every effect the monitor compiled is compiled
+again. And CI no longer sees it: the suites that took past five minutes
+on WARP now return early and count as passed (740 run, 740 passed, run
+38302292964), their "skipping" lines swallowed by `store-success-output =
+false` (`nextest.toml:31`). Windows is the platform with the most
+downloads. Share compiled pipelines between siblings; check cancel between
+pipeline builds and show a "preparing effects" stage; make a skip visible
+in the step summary; run the full suites on a schedule with
+`CONCAT_REQUIRE_GPU=1`; say in the 0.2.7 notes that the compile cost
+remains. (verified against the runs and the code; the minutes are from
+#297 and be1b882's message)
+
+### Medium
+
+**5. More than 24 titles drop out of the monitor during a drag, and every
+title is re-treated each frame.** While a gesture is live
+(`begin_echo`, `studio.rs:2474`) the monitor paints every title live and
+hands each to `monitor.hold_still` (`studio.rs:2600-2618` →
+`preview.rs:152-154` → `pool.rs:628-652`); the pool keeps held stills
+FIFO, capped at `STILLS_HELD = 24` (`pool.rs:541, 634-638`). Each pass
+re-inserts in clip order, so only the last 24 survive; a title not held is
+looked for as a file at `memory://titles/…`, fails, and is dropped without
+a word (`concat-export/src/lib.rs:1913-1945`). Auto-captions add clips in
+time order, so on a captioned timeline the early captions vanish while
+anything is dragged, including a caption whose own size is being dragged.
+And `hold_still` calls `forget` for every title every pass (`pool.rs:640,
+675-685`), even for the same `Arc`, so each is treated again: the
+gesture half of #261 is undone. Hold only the titles the plan at the
+playhead covers, or stamp held stills per pass as `LivePaintings` does;
+skip `forget` when the frame is the same. (verified by reading)
+
+**6. Two saves of one project write the same scratch file.**
+`projects::save` always writes `concat.json.saving` (`concat-host/src/
+projects.rs:174`) and renames it. The window saves on a pool of three
+workers (`concat/src/host.rs:146-190`; `studio.rs:2500-2518`) and
+`close_project` saves on the UI thread (`studio.rs:7230-7235`): Ctrl+S
+twice, Ctrl+S during an autosave, or a close during a save puts two
+writers on one path. Interleaved writes can rename a manifest that does
+not parse ("not a Concat project", `projects.rs:224-226`); the loser's
+rename fails and toasts "could not save", and a close aborts. Take a
+per-path lock, give each save its own scratch name, and skip a save older
+than the one on disk. (mechanism verified; the corrupting interleaving
+reasoned)
+
+**7. On OpenGL ES, the 30-second device wait is about 2.15 seconds.**
+`wait()` passes `Some(DEVICE_WAIT)` (`gpu.rs:1311-1318`); wgpu-hal's GLES
+backend clamps it to `u32` and then `i32::MAX` nanoseconds
+(`wgpu-hal-30.0.1/src/gles/device.rs:1757-1764`, `gles/fence.rs:
+159-163`). On a GLES phone, the #312 device, a frame slower than that
+fails `read_back`, kills the compositor, is retried on a fresh device,
+asks for a software adapter a phone does not have, and ends the export as
+a lost GPU (`concat-export/src/lib.rs:1245-1247`). Not a regression (the
+old indefinite wait was capped the same way), but the doc
+(`gpu.rs:183-191`) is wrong on the platform it was written for. Loop
+short waits until `DEVICE_WAIT` has passed. (reasoned from wgpu's source)
+
+**8. The captions sheet can transcribe a clip other than the one it
+shows.** `Select`, `SegmentedControl` and `Switch` assign their own value
+(`select.slint:127`, `segmented-control.slint:43`,
+`inspector/controls.slint:559`), which in Slint drops the parent's
+binding. The sheet binds Switch `checked: root.from-voice`
+(`captions.slint:109`), Audio (`:123`), Model (`:134`), Placement (`:207`)
+and Size (`:216`) that way, is only hidden when closed
+(`modal.slint:34`), and Rust resets the fields on every open
+(`captions.rs:201-221`). So: Generate captions on clip A, pick B in
+Audio, close, Generate captions on clip C: the list shows B, Rust
+transcribes C. Placement and size have shown stale values since before
+this week. 590e483 and f1ccdfe fixed this class at two sites a day
+earlier; about 33 such bindings remain in `ui/`. Make the three
+primitives controlled (emit, never assign), which fixes every site at
+once; the sheet's own script box shows the pattern (`captions.slint:
+72-86`). (verified in code; the behaviour reasoned)
+
+**9. An HEVC or AV1 export still opens NVENC first, so #309 moved from
+launch to export.** b4adc57 stops start-up from opening a vendor encoder
+(`encode.rs:171-180, 442-479`, with tests), but the export always asks for
+hardware (`hardware: true`, `concat-export/src/lib.rs:1354`;
+`ExportRequest` has no field for it), and `encoders(true)` leads with
+`hevc_nvenc` or `av1_nvenc` (`encode.rs:117-130`). On the Optimus laptops
+#309 describes, `cuCtxCreate` segfaults at the first such export; the way
+out is `CONCAT_HARDWARE_ENCODE=0`, which nothing in the app mentions.
+Write a sentinel before a vendor-driver open and treat it as refused at
+the next launch if it is still there; expose a "Hardware encoding"
+setting. (verified by reading)
+
+**10. Any validation error now reads as a lost GPU.** A Validation scope
+error marks the compositor dead (`gpu.rs:3410-3437`); the export opens a
+second device, then WARP, draws the same frame on each, and ends "the GPU
+device was lost 3 times part way through the export - reset, out of
+memory, or not answering for 30 seconds" (`concat-export/src/lib.rs:
+1231-1243`). For a deterministic refusal that sends the person to their
+driver. On Windows it also opens the second device that `render_on`'s own
+doc (`:874-880`, #202) records as dying in NVIDIA's driver on laptops.
+Separate refused from lost: a validation error fails the export with
+wgpu's text; only loss, out-of-memory and the timeout recover. (verified
+by reading)
+
+**11. The 11th's finding 8, the half not closed.** `combine`, which
+packaged transitions use, has no error scope and the counter is not read
+around it (`gpu.rs:3450-3518`); a pass whose pipeline the driver refuses
+is filed in `refused` and skipped for good (`:2552-2558, 3056-3059`), so
+the exported file lacks the effect and only the log says so; and the
+window's lent export compositor carries a counter nothing increments
+(`drawing_in` makes a fresh one, `gpu.rs:1291`; the window's handler only
+logs, `concat/src/gpu.rs:113-115`). Scope `combine`; fail an export whose
+effect could not be drawn; share the counter. (verified by reading)
+
+**12. A phone that cannot start just vanishes, and an iPhone's work cannot
+leave the app.** `report_startup_failure` is a no-op on Android and iOS
+(`platform.rs:459-460`); at `HEAD` the end of `start()` exits 0 whatever
+`app.run()` returned (`lib.rs:1928-1932`), and `android_main` logs and
+returns (`concat-android/src/lib.rs:395-397`). The uncommitted `leave(1)`
+gives CI an exit code, not the person a message. Separately,
+`panes/start.rs:94-95, 116-118` makes Documents/Concat "so that the Files
+app shows it", but `src/scripts/ios-app.sh:54-97` sets neither
+`UIFileSharingEnabled` nor `LSSupportsOpeningDocumentsInPlace`, so Files
+does not. Show a Toast or dialog through the bridge on a failed start;
+add the two plist keys. (verified)
+
+**13. `locales.py --check` passes with keys missing.** A stale key sets
+`failed` (`scripts/locales.py:242-244`); a missing one is printed only
+(`:245-246`), against the docstring's "exit 1 on anything out of step"
+(`:5-7`). With 25 locales, three from outside contributors, a key added
+to `en.json` alone ships in English in 24 languages with CI green
+(`ci.yml:141`), and `i18n.rs:243-247` falls back silently. Fail on
+missing; add a placeholder check. (verified)
+
+**14. The release notes carry CI internals and a claim the next commit
+reverts.** `release-notes.sh:43` drops `ci:`, `docs:` and `chore:`, not
+`fix(release):`, so "The Windows smoke test and its emulated Ivy Bridge
+are two steps…", "The Linux and macOS smoke scripts can run…" and "A
+closing app ends with _exit, so v0.2.7-rc.2's arm64 bundle…" become 🐛
+lines; be1b882's "a blur or a glow no longer stalls an export" sits beside
+3178b78's "no DXC in 0.2.7"; and `i18n:` subjects keep their prefix
+because the type pattern `[a-z]+` (`:33, 53`) has no digits. Filter the
+`release` and `ci` scopes and `i18n`, and edit the 0.2.7 draft by hand.
+(verified by running)
+
+**15. Windows ships the CPU ONNX Runtime; concat-speech still asks for,
+logs and retries DirectML.** `build-app.yml:381-395` stages Microsoft's
+CPU `onnxruntime.dll` (a deliberate trade for old CPUs);
+`concat-speech/src/lib.rs:27-41` keeps `accelerated()` true on Windows,
+`chatterbox.rs:303-314` registers DirectML (silently dropped) and logs
+"asked to run on DirectML". So e5d8197's retry treats a real CPU failure
+on Windows as an accelerator failure, loads the model again and reads the
+whole text twice before reporting it, and every Windows diagnostic names
+a provider that is not there. `ARCHITECTURE.md:458` and
+`concat-vision/src/runtime.rs:7` say the same. Make `accelerated()`
+false on Windows until a DirectML runtime ships, and say "CPU" in the
+log. (verified by reading)
+
+**16. Cutouts and Chatterbox on an Intel Mac probably fail at their first
+session.** The workflow says so itself (`build-app.yml:91-96`): ort
+2.0.0-rc.13 asks for API 27, the bundled `onnxruntime 1.23.2` serves 23,
+"the first ONNX session … should fail". `--smoke` opens no session, and
+`macos-x86_64` has never finished a release build. An older `api-*`
+feature for that target, or the ONNX features hidden there, and a smoke
+flag that opens one small session. (reasoned from the workflow's note)
+
+**17. `--smoke` stops before the window exists, and the scripts say it
+draws.** It exits after `publish`, before `app.run()`
+(`concat/src/lib.rs:1900-1907`), as its own doc says (`:145-157`);
+`smoke-linux.sh:10-11`, `smoke-macos.sh:10-11`, `smoke-windows.ps1:17-18`
+and `build-app.yml:32-34` say "through the first frame of the window".
+Window and surface creation, the Skia renderer, the swapchain and the
+event-loop errors of #312 are never exercised by any launch check. Run
+the loop in smoke mode and quit from a one-shot timer after a few frames;
+the runners have Xvfb and WARP for it. (verified)
+
+**18. The instruction-set guard cannot catch a repeat of #193.**
+`isa-allow.txt:62-63` lets `concat` and `concat.exe` carry any number of
+AVX2, FMA, mulx/rorx and AVX-512 instructions, BMI only warns, and the
+"PROVISIONAL… set the limit once it has run" step (`:56-61`) was not
+taken after rc.2 printed real counts (avx2 18 127, fma 3 405, bmi 21).
+The emulated Ivy Bridge runs `--version` only (`smoke-windows.ps1:158`,
+`smoke-linux.sh:106`), which never enters ggml. So whisper's CPU backend
+going back to AVX2 would pass both. Set the limits from the measured
+counts with a margin, and run a short transcription under the emulator.
+(verified from the rc.2 log and the files)
+
+**19. The bundle workflow can only be run by pushing a tag.**
+`build-app.yml:45-74` has `workflow_call` only, and `release.yml` is its
+one caller: rc.1 found three script bugs, rc.2 one crash, and two fixes
+since are unproven. Add `workflow_dispatch` with a `ref`, or a nightly
+without `publish`. (verified)
+
+**20. `CODE_SIGNING.md` describes a different process from the
+workflow.** It says FFmpeg's and sherpa-onnx's DLLs are not signed with
+the certificate (`CODE_SIGNING.md:12-16`) and that every signing request
+is approved by hand (`:27`); the workflow signs every `.exe` and `.dll`
+not already validly signed, deliberately, because Smart App Control
+refuses unsigned DLLs (`build-app.yml:913-962`; #306, #311), from a PFX
+secret, automatically. One of them must change before SignPath reads the
+policy; the Smart App Control reason argues for the workflow. (verified)
+
+**21. Dependabot's cargo group is red by design.** `dependabot.yml:27-36`
+groups every crate; #315 moves `sherpa-onnx` 1.13.7 to 1.13.8, and the
+guard that `deps.lock` matches `Cargo.lock` (`ci.yml:125-128, 326-329`;
+`build-app.yml:433-436, 473`) fails every job, holding the other 16
+updates with it. Ignore `sherpa-onnx*` in the group and bump it by hand
+with `deps.lock`. (verified against run 38295637177)
+
+### Low
+
+**22. The 4 px drag threshold is under a finger's slop.** `lanes.slint:
+591-592` treats a press as a drag past 4 logical pixels; Android's own
+tap tolerance is 8 dp. A tap that wobbles 5 to 7 px on a phone still
+moves a clip: #247 survives on the platform with the most downloads. 8
+to 10 px when `platform::phone()`. (reasoned)
+
+**23. #151 comes back after Browse.** `free_untitled` runs on `Compose`
+and on an empty name at `Create` (`panes/start.rs:160-166, 243-247`), not
+on `FolderChosen` or `LocationEdited` (`:215, 175`); browsing to a folder
+that already holds "Untitled project" and pressing Create says it exists.
+(verified)
+
+**24. Detach on a video muted on purpose unmutes it.** `detach_plan`
+answers `Again` for any muted video with no detached sound
+(`studio.rs:10446-10447`), and `detach_audio` batches `muted:
+Some(false)` with the detach (`:10293-10305`); a mute the person chose
+cannot be told from one a deleted detach left. Mark the detach's mute, or
+mute the new sound clip in the `Again` case. (verified)
+
+**25. Any argument but the three flags refuses to start.**
+`launch::parse` returns the usage for anything else (`launch.rs:43-47`),
+which `main.rs:31-33` shows as "Concat could not start" in a dialog.
+Before 45c061f the arguments were ignored. The Linux packages' desktop
+entry is `Exec=/opt/concat/concat %U` (`build-app.yml:803`), and Windows'
+"Open with" passes a path: either now opens an error box instead of the
+app. Log and ignore unknown positionals, or drop `%U`. (verified)
+
+**26. Ten commits say "Fixes #N", against the repository's rule, and the
+push closed the issues.** 132f68d (five issues), 590e483, f1ccdfe,
+e5d8197 and e1b64a4 (two) use "Fixes"; GitHub closed #299 on the push at
+04:52 and it was reopened by hand at 04:54 (the issue's events). All are
+open now. "Refs #N" in the body, as the rule says.
+
+**27. Right-to-left languages are laid out left to right.** Arabic,
+Hebrew, Urdu and Persian ship with faces that shape them (`fonts.rs:53-58,
+104-112`), but nothing in `ui/` mirrors a row or flips an alignment. A
+Slint limit; say so in the README until it is handled.
+
+**28. The CI concurrency group cancels pending runs on `main`.**
+`cancel-in-progress: false` still cancels a pending run when a newer one
+queues; seven pushes have no CI record (d118484, 049823f, 9737836,
+0a61b65, 2d3e651, 0714fb8, 91a6df0), against the comment at
+`ci.yml:26-32`. Key the group on `github.sha` for pushes.
+
+**29. `ubuntu-latest` becomes Ubuntu 26 from 19 October.** Every run
+carries the annotation. Jobs on it: `ci.yml:37` (with apt names that may
+move), `mobile.yml:92` and the new emulator job, `release.yml:42, 118`,
+`nix.yml:35`, `models.yml:41`. Pin `ubuntu-24.04` before then.
+
+**30. The live-title ceiling is a constant.** `LIVE_CEILING` is 512 MiB
+(`titles.rs:98`), not sized from `memory::total()` (`memory.rs:10-14`)
+as the frame pool is (`pool.rs:577`); on a 3 to 4 GB Android phone a
+captioned timeline mid-gesture can hold that much on top of the pool.
+
+**31. The 8-bit fallback clips HDR with a log line.** On a phone that
+falls back to `Rgba8UnormSrgb` (`gpu.rs:174-181, 784-790`), an HDR
+timeline exports clipped and tagged HDR. Refuse HDR there, or say so in
+the sheet. (reasoned)
+
+**32. The API, the server and the CLI.** The `gpu` capability can never
+appear (`concat-api/src/lib.rs:180, 370-372`; the test pins its absence,
+`:1066`); a gRPC `Events` stream and `Server::stop` can wait on each
+other (`grpc.rs:167-170`, `lib.rs:281-293`; CLI only); `concat-cli
+render` writes in place with no `MAX_SIDE` check (`main.rs:417-426`).
+
+**33. Doc drift, continued.** Three doc comments stacked or misattached
+(`studio.rs:2127-2138`, `:7556-7565`, `:10258-10265`, the last leaving
+`clip_action` at `:10326` with none); `DEVICE_WAIT`'s doc (finding 7);
+two contradictory paragraphs on `read_back` (`gpu.rs:3312-3321`);
+`Gpu::acquire` drops a `request_device` error unlogged, so "no GPU
+adapter" is printed for a device failure too (`concat/src/gpu.rs:
+96-100`); `dirs.rs:32` ("downloaded models", now logs, titles, cards and
+updates too); `README.md:19` (a 0.2.7 badge whose alt text says 0.2.5,
+before 0.2.7 is out), `:52` (no iOS in About); `models.yml:19-20` against
+`:87`; `ARCHITECTURE.md`'s known issues (`:652-665`) stale since the day
+after they were written.
+
+**34. Pull-request hygiene.** #304 is still open although it landed as
+c0c8827; Dependabot's #314 failed on the old Windows hang and needs a
+re-run, and its `actions/cache` bump will conflict with the uncommitted
+`mobile.yml`; eleven contributor PRs (#262 to #275) are six days
+unreviewed.
+
+---
+
+## 7. The uncommitted work, before it is committed
+
+Reviewed on the working tree, because once committed it is in the
+release.
+
+**The encoder retry (`concat-media/src/encode.rs`).** Right by FFmpeg's
+contract: on "try again", read the packets waiting and send the same
+frame again (`patiently`, `send`), for the frame and for the end of the
+stream; software encoders never answer it, so the desktop is unchanged;
+`written` still counts only sent frames; a stall leaves no file, because
+the scratch is removed and the output never touched. A good pure test
+with a fake. Before committing:
+- The old line `/// Writes every packet the encoder has ready.` now heads
+ `send`'s doc, and `drain` has none. Move it back.
+- After the end of the stream, `drain` still stops at the first "try
+ again" (`encode.rs:1341` in the tree). The contract says it cannot
+ happen there; this change exists because MediaCodec's wrapper breaks the
+ contract on the sending side, and if it does on the receiving side the
+ tail of every phone export is lost with no error. Make the final drain
+ patient too, in the same commit.
+- "Refs #286", not "Fixes": #286 says only "missing codecs", and no
+ export has been seen working on a phone.
+- The Android launch check never exports. A one-second 720p H.264 export
+ in it would exercise `h264_mediacodec` on the emulator's
+ `c2.android.avc.encoder`.
+
+**The phone launch checks (`mobile.yml`, `smoke-android.sh`,
+`smoke-ios.sh`).** Well designed: Android's fatal lines are matched only
+for the app's own process or package, the foreground activity must be the
+app's, the native first line must appear, and logs and a screenshot are
+kept on every outcome; iOS checks the binary's platform and uses a fresh
+simulator, deleted in a trap. Before committing:
+- Run it once by `workflow_dispatch` on `main`; it gates `publish`
+ (finding 3).
+- `smoke-ios.sh:184-188` passes on "smoke: ok" and ignores the exit
+ status, so a crash after the line (rc.2's Linux failure exactly) passes
+ on iOS.
+- `smoke-android.sh` has no `trap`; a `set -e` exit at `getprop` or
+ `unzip` (`:128-136`) leaves `adb logcat` running and keeps no logs.
+- The simulator's environment is written inline in `mobile.yml`;
+ `src/scripts/mobile-env.sh`, "the one description of the build
+ environment", knows only the two device targets.
+- Check `git ls-files -s scripts/ci/smoke-*.sh` shows `100755` after
+ `git add`; b161c1d exists because it did not once.
+- An emulator is not a phone: SwiftShader under ARM translation can pass
+ what fails on a Mali Android 11 (#312). Record its first result as the
+ baseline.
+
+**The failed-loop exit (`concat/src/lib.rs`).** Logging the error and
+exiting 1 through `leave` is right and fits the phone checks' patterns;
+it is for CI, not the person (finding 12).
+
+---
+
+## 8. What is good, and should stay that way
+
+- **The 11th's first recommendation worked the first time.** nextest
+ with a per-test timeout (`src/.config/nextest.toml`) turned a
+ ninety-minute silent hang into four named tests on its first Windows
+ run, and the JUnit report is kept on failure.
+- **The supply chain is one file.** `scripts/ci/deps.lock` pins 24
+ downloads by digest, `fetch.sh` refuses and deletes a mismatch and
+ prints both digests (`:66-78`), FFmpeg is a dated build, and a guard
+ holds sherpa's pinned file to `Cargo.lock`'s version. Every action is
+ pinned by SHA with its tag in a comment.
+- **A release cannot publish what did not start.** Every desktop bundle
+ is launched where it was built before upload; Windows' launch audits
+ the bundle's imports against a must-ship list and cuts `PATH` to
+ System32 (`smoke-windows.ps1:96-150`); Linux installs the `.deb` and
+ runs the AppImage in clean Ubuntu and Debian containers and checks that
+ start-up opens no CUDA library (`smoke-linux.sh:85-150`, #309); the
+ release is a draft that needs every bundle and both phones.
+- **Platform rules are pure functions of their inputs.**
+ `dirs::convention` and `apple` (`dirs.rs:140-161`),
+ `updates::windows_kind` (`:286-294`), `start::location_on`,
+ `encodable_among` and `leads_with_hardware` with injected `opens` and
+ environment (`encode.rs:442-479`): each platform's rule is tested on
+ every runner, and b4adc57's tests panic if anything is opened.
+- **Every native dialog has one seam.** `after_dialog` and
+ `pick_files_then` (`host.rs:352-400`): a dialog runs with nothing
+ borrowed, and the phone and the desk share the call.
+- **Fonts are done properly.** Subset Noto faces with source and SHA-256
+ per face (`fonts.rs:44-80`), each fallback once, and a test that every
+ character of every shipped locale has a glyph.
+- **The phone GPU path is coherent.** `limits_for` asks for the adapter's
+ own limits where it is below WebGPU's (`gpu.rs:548-552`), the 8-bit
+ working format follows into every pool, target and byte count, and the
+ fallback is tested against the half-float picture to ±2 levels
+ (`gpu/tests.rs:1825-1847`).
+- **Device loss is finally tested end to end** through `render_on` with a
+ dead lent compositor, redrawing the frame rather than dropping it
+ (`concat-export/src/lib.rs:1517-1541, 2640-2706`).
+- **`LivePaintings` is the right shape** (`titles.rs:116-190`): one
+ policy, stamped per pass, a pure test and a test that fails under the
+ old FIFO. The pool's held stills should copy it.
+- **Honest reverts and honest messages.** 3178b78 says what failed and
+ ships the known state; b161c1d, 2d3a43e and c06560e name the release
+ candidate that found each bug, and `leave()`'s doc says what is guessed.
+- **The 70 `unsafe` sites each carry a SAFETY line**, and concat-core is
+ still free of `unsafe`.
+
+---
+
+## 9. Where we stand
+
+**Can 0.2.7 ship?** Not from today's tree. `main` is green, but no
+release run has passed every desktop bundle's launch check, the phones'
+new checks have never run, and the final release's notes would be three
+lines. What stands between the tree and a tag, in order:
+
+1. Fix the previous-tag lookup in `release.yml:170` (finding 1).
+2. Finish the uncommitted work: the doc line, a patient final drain,
+ the two smoke-script fixes, "Refs #286"; commit it on its own (section 7).
+3. Run `mobile.yml` by hand on `main` and look at its result (finding 3).
+4. Tag `v0.2.7-rc.3` and let every target run to the end; c06560e and
+ 2d3a43e are proven there or not at all.
+5. If green, tag `v0.2.7`, and edit the draft's notes by hand before
+ publishing: drop the CI-internal lines and the DXC pair, and say that a
+ blur or glow export on Windows can still take minutes to start
+ (findings 4, 14).
+
+The 11 October audit ended with nine recommended steps. One day later:
+
+| Step (11 Oct) | Landed |
+|---|---|
+| 1. Make the Windows job name its hang; skip GPU suites on a software adapter; engine timeout; pin the toolchain action; tag 0.2.7 from a green `main` | all of it but the tag: named, skipped, timed out, pinned; two candidates tagged and failed |
+| 2. The three remaining dialogs; relink, cache clear and sweep on workers; trials off the UI thread | the dialogs (132f68d); nothing else |
+| 3. The paste: trimmed placement, the full patch for titles, clear the clipboard; move placement into concat-project | nothing |
+| 4. `Server::stop`, `Builder::spawn`, `catch_unwind`, a protocol whitelist, a write timeout, `exists` and `is_sane` on the API | nothing |
+| 5. Count uncaptured errors; scratch-name suffix; one past-the-end rule; a VFR fixture | the counter and scopes (e1b64a4), with half left (finding 11); nothing else |
+| 6. Digests, actions by SHA, `inputs.tag` through `env:`, Windows signing secrets, a release with the real Android key | all of it for the desktop (d118484 and after); the phones' downloads remain |
+| 7. Pinned-frame budget, WAV header, group clamp, speed-curve ripple, `DOCUMENT_VERSION` | nothing |
+| 8. The literals through `I18n`; a locale check that flags English and prints anywhere | the check prints and counts English (9737836); the literals as they were, and two new ones |
+| 9. The 28 September list; Windows text and the phone's export on real hardware | nothing seen |
+
+Three and a half of nine in one day, and the two that mattered most for a
+release (the gate and the supply chain) in full. The day's other work was
+the tracker, and it was taken well: every bug fix landed with a pure test
+that fails without it. The pattern the 11th named is still there, a
+little smaller: a fix that claims a little more than it does (the
+Windows suite could finish; a blur no longer stalls; #261's titles stay
+painted; "Fixes") and is caught by the next reading or the next tag. The
+new risk is the opposite of last week's: CI is green because the tests
+that see Windows' problem no longer run there, and the release pipeline
+can be exercised only by spending a tag.
+
+**Recommended order.** Each line is an hour to a few days, and each
+closes a finding above.
+
+1. The five release steps above. (1, 3, 4's notes, 14; section 7)
+2. iPhone import through a document picker, or a notice until it exists;
+ a failed start shown on a phone; the two plist keys. (2, 12)
+3. The captions sheet: make `Select`, `SegmentedControl` and `Switch`
+ controlled; then the ~33 bindings are right everywhere. (8)
+4. The pool's held stills stamped per pass and `forget` skipped for the
+ same frame; one scratch name per save with a per-path lock. (5, 6)
+5. Windows shaders: share compiled pipelines between the monitor and the
+ export; cancel between pipeline builds; a visible skip and a scheduled
+ `CONCAT_REQUIRE_GPU` job; then try DXC again on real hardware. (4)
+6. The GPU's two outcomes, refused and lost; `combine` scoped; a refused
+ effect fails the export; patient short waits on GLES; a sentinel and a
+ setting for vendor encoders. (7, 9, 10, 11)
+7. CI: `push` on `mobile.yml`; `workflow_dispatch` on `build-app.yml`;
+ `--smoke` through a few real frames; instruction limits from rc.2's
+ counts and a transcription under the emulator; pin `ubuntu-24.04`
+ before 19 October; the concurrency key; Dependabot's sherpa exclusion;
+ `locales.py` failing on missing keys. (13, 17, 18, 19, 21, 28, 29)
+8. Windows' ONNX Runtime honest in speech and vision; the Intel Mac's
+ runtime matched to ort or the features hidden; `CODE_SIGNING.md`
+ brought in line with the workflow. (15, 16, 20)
+9. Then the 11th's steps 3, 4 and 7 as they stand (the paste, the
+ server, the pinned-frame budget and the rest), and the 28 September
+ list.
+
+---
+
+## 10. Caveats
+
+Scores are judgements over code, calibrated against the 11 October scores
+but not measured against users or another editor. Every crate was read by
+one of four reviewers; the headline findings were re-read by hand.
+Findings marked (verified) were read at the line and, where a command
+could show it, run; (reasoned) means the path was followed through the
+code, or through wgpu's source, and not executed. The checks ran on an
+Apple silicon Mac with Metal, which is neither Windows' WARP nor a phone:
+the Windows stall, the GLES wait and the iPhone's import are read from the
+code, the workflow runs and the issues, not reproduced. The test run
+included the uncommitted encoder change. The app was not launched; no
+release bundle or phone build was made here; `cargo audit` was not run.
+The Linux arm64 exit fix (c06560e) compiles for that target and is
+explained in its commit, but it has not yet run on the runner that failed.