Skip to content

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

docs(audit): the 15 October audit

Jareerdcf6384b71a5
diff --git a/audit/2026-10-15/host-server-api-cli.md b/audit/2026-10-15/host-server-api-cli.md
new file mode 100644
index 00000000..e5685e05
--- /dev/null
+++ b/audit/2026-10-15/host-server-api-cli.md
@@ -0,0 +1,74 @@
+## host, server, api, cli: reviewer 2, read at ab21457 (main moved to 2e4c4d5 during; only 8 lines of concat-host/tests/export.rs touched)
+Verified = read at line; reasoned/suspected otherwise. No cargo run.
+
+### Re-verified
+- 12Oct#5 pool 24 stills FIFO, forget each pass: UNCHANGED pool.rs:541, 628-652, 675-685; hold_still now concat-host/src/preview.rs:169.
+- 12Oct#6 shared save scratch: UNCHANGED, worse outcome silent. projects.rs:174 fixed concat.json.saving; studio.rs:2633-2650 saves on 3-worker pool, dirty=false (:2638) before write; two renames land either order -> older doc on disk, no error, no dirty flag. Close saves on UI thread (:7300-7304). (reasoned)
+- 12Oct#30 live-title ceiling constant: titles.rs:98 512 MiB.
+- 12Oct#32 api/server/cli smalls: gpu capability never appears (concat-api/src/lib.rs:180, 370-372, test pins absence); stop waits on open gRPC Events (grpc.rs:162-170, lib.rs:281-296); CLI render in place, no MAX_SIDE (concat-cli/src/main.rs:419-431).
+- 12Oct#33 dirs.rs:32 "downloaded models": unchanged.
+- 11Oct#21 MCP/CPU fallback docs: PARTLY. README fixed; concat-api/src/lib.rs:7 still "an MCP server"; concat-host/Cargo.toml:37-38 CPU compositor claim.
+- 11Oct#6 Server::stop hangs on wildcard bind on Windows: UNCHANGED concat-server/src/lib.rs:270-272, 281-283; on UI thread settings.rs:557-558.
+- 11Oct#7 OS-refused thread kills accept loop: UNCHANGED json.rs:128, 183, :305. panic=abort in [profile.app] -> aborts whole app.
+- 11Oct#10 API writes unchecked: export checks only roots (api lib.rs:520); create no size/rate check (:380-397); projects::create writes rate_den 0 (projects.rs:106-148) which open refuses (:229-231), folder left.
+- 11Oct#15 server edges: grpc.rs:74-76; poisoned subscriber list silences events (hub.rs:54-57); hub.rs:155-157; CLI api ends on non-UTF-8 line (main.rs:175-176); doc on impl Drop (api lib.rs:868-870), key_of (:933) none.
+- 11Oct#16 stop blocks window; jobs/workers die silently: lib.rs:293-296 waits api.finish() joining export (api lib.rs:748-752); catch_unwind (hub.rs:91-104) useless under abort; ExportJob::run (:761-786) no catch -> no export.failed; playback.rs:683, 693 .unwrap() on locks.
+- 11Oct#17 reads unconfined; caller strings reach FFmpeg: api lib.rs:226-230, 326-328, 465-469.
+- 11Oct#23 host lows: silent data-folder move dirs.rs:108-119; any Program Files program = MSI install updates.rs:286-302, case-sensitive starts_with; templates.rs:356 poster as preview.jpg.
+- 11Oct#25 WAV header overflow record.rs:287, 315.
+- 11Oct#29 --socket deletes whatever file: concat-server/src/lib.rs:195.
+- S1 update trusts digest beside package: digest from same manifest as URL (updates.rs:416-428, 465), no signature; setup.exe no Authenticode (:491-495); hdiutil -noverify (:539); xattr -dr quarantine (:584-586) bypasses Gatekeeper; any release >=0.2.5 installable incl. downgrades.
+- S2 API root is home: concat-server/src/lib.rs:80-86.
+- S3 fd per connection, seats before auth, accept spins: UNCHANGED, worse than scored (finding 1).
+- S4 Clear cache deletes reversed/enhanced media: reverse.rs:88, enhance.rs:71, projects.rs:341-350.
+- S6 embedded API own GPU device: api lib.rs:179; Vulkan window export too now.
+- S8 HDR proxy name/contents disagree: unchanged.
+- S9 API composites on hub thread; slow subscribers dropped silently: api lib.rs:678-706; json.rs:319-326; grpc.rs:74-76.
+- S13 mixer scans every clip per sample: playback.rs:1139-1166.
+- S14 missing font silent: titles.rs:441.
+- S16/S17: connections() counts subscribers (lib.rs:257-259); Unix socket umask perms (:196).
+Tally ~25: 0 fixed, 1 partly, 1 escalated, rest unchanged. None of the 11 Oct "server step" items landed.
+
+### Commits
+- 14b5998: does it. off_thread (preview.rs:146-153) None on Vulkan; export (panes/export.rs:518), cards (studio.rs:7481), merge (studio/consolidate.rs:116); look trials stay sibling (studio.rs:7543). Diagnosis right (Skia submits to raw VkQueue without wgpu lock). Misses: Android also Vulkan (concat/src/gpu.rs:52-57), commit titled linux doesn't say; no test. Refs #324.
+- 8f5b141 crash line + machine header (#326): Good: one line with code, thread, module+offset; chains to previous handler/filter; SA_ONSTACK, CRASHING guard, dup'd fd; installed before --version (main.rs:19), open_logging (lib.rs:106), CLI (main.rs:128); child-process test passed Linux/macOS/windows-2022 (run 38435821575, job 115429378305); machine.rs writes build, commit, OS, CPU name+flags. Problems: dladdr (crash.rs:217) takes loader lock (glibc dl_load_lock, bionic g_dl_mutex, dyld): blocks if another thread holds it; SAFETY "only reads loader's tables" (:213-214) false. to_string_lossy (:218) can allocate. Windows: std::thread::current() (:398) allocates on non-Rust threads (ORT pool, drivers, FFmpeg) and panics after TLS teardown -> fail-fast abort; GetModuleFileNameW (:442) loader lock. ~3 KB buffers on faulting stack, no alt stack on Windows (suspected stack-overflow loss). Crash during crash: Unix correct; Windows nested exception loses line (write after describe :379-382). Handles dup'd once, CLOEXEC/non-inheritable, swapped atomically; keep closes old fd (:137-141) race only if log opened twice. Handler re-pointed before logger install can fail (logs.rs:78 vs :89). Untested: write to log file; crash on non-Rust thread.
+- fc7e6a9 consolidate (host): consolidate.rs 219 lines captures clips/media/settings, rechecks before adopting (:52-68); normal exporter with effect_time_offset; pads to whole frame; window deletes unadopted render; 2 unit + 2 end-to-end tests (+284 tests/export.rs) render, probe, adopt, export again, undo, redo, cancel: best-tested work in period. Misses: API exposes raw command (6); CRF16 "fast" hard-coded, HDR -> 10-bit HEVC hardware-first (consolidate.rs:120-132, NVENC on Optimus); doesn't wait for cutout masks (9); adopted-then-undone renders never collected (14); no Refs.
+
+### New findings
+HIGH: none.
+MEDIUM
+1. Any process can exhaust the window's fds via Remote API without a token (verified). Every TCP/Unix connection leaves a dup'd socket in connections (json.rs:109-122, 164-177) until stop; seat before handshake (:102, :157). A few hundred connections (256 macOS, 1024 Linux): saves fail "Too many open files" (projects.rs:186-187), decoders fail, accept loop spins 100% CPU (:100, :156). A script connecting per call hits it in normal use. Windows: handles grow. Remote API off by default. Fix: deregister closer when thread ends; seat after auth; back off on accept error.
+2. Crash handler takes locks and allocates despite doc (crash.rs:30-33) and commit (reasoned). Crash while another thread loads a library (ORT providers, Vulkan drivers, FFmpeg hw libs: #326's window) can hang with nothing logged. Windows heap fault on non-Rust thread can refault in filter -> line lost. Crash during thread exit -> fail-fast attributed to concat.exe. Fix: write raw code + PC first; record module ranges at install; drop thread::current() from filter.
+3. Android: every export, merge, card round opens a second Vulkan device (path verified, impact suspected). Android Vulkan (concat/src/gpu.rs:52-57) -> off_thread None (preview.rs:152); export/merge own instance+device; cards WgpuCompositor::new() every round (cards.rs:404-406). 3-4 GB phones, #312-class drivers. Fix: check Skia on Android (Vulkan or GL); if GL keep sibling().
+LOW
+4. api lib.rs:7 MCP.
+5. Server::connections counts subscribers; stale "N connected".
+6. API ConsolidateClips adopts any file (commands/clips.rs:612-645); AddMedia returns early if path in bin (commands/media.rs:19-25); no render-a-selection verb. Hide or add edit.consolidate.
+7. Weak tokens accepted (settings.rs:575), no auth throttle (json.rs:267-291), 0.0.0.0 binds allowed (tests settings.rs:983, 996); LAN brute force -> writes anywhere under home, e.g. overwrite a file with PNG via preview.frame (api lib.rs:642-656).
+8. Two crash codes can never arrive: STATUS_STACK_BUFFER_OVERRUN, STATUS_HEAP_CORRUPTION (crash.rs:358-359): fail-fast skips the filter.
+9. Merge may bake a clip whose cutout masks aren't analysed (suspected); API export awaits masks (api lib.rs:574-580).
+10. One-channel output loses right channel (playback.rs:1172-1175): HFP Bluetooth.
+11. Audio callback frees memory on RT thread (playback.rs:1107).
+12. README advertises gRPC API and CLI no download ships: releases build only -p concat (build-app.yml:523) without grpc feature (concat/Cargo.toml:133-134), no concat-cli shipped; window serves JSON-RPC over TCP only.
+13. Update/model downloads no size limit (models.rs:258-276). Suspected: projects.rs:194-195 single rename no retry under OneDrive/AV on Windows -> "could not save".
+14. Merged renders never collected (studio/consolidate.rs:102-109); undo keeps file; Clear cache spares media/.
+
+### Numbers
+host 11 971 (+1 129) tests 99 (22) +10, unsafe 42 (+38: crash.rs 32, machine.rs 6), unwrap 286 (2) (+37 all tests), panic 3 (0)
+server 1 393 (0) tests 12 (4) unsafe 0 unwrap 28 (0)
+api 2 352 (0) tests 21 (2) unsafe 0 unwrap 23 (0)
+cli 517 (+3) tests 3 (1) unsafe 0 unwrap 2 (2)
+Biggest files: playback.rs 1 347, api lib.rs 1 481, media.rs 1 093, titles.rs 878, updates.rs 862, cards.rs 653, templates.rs 618, crash.rs 556, json.rs 508.
+Biggest fns: brush read 154, playback stream_for 147, cutout analyse 140, reverse footage 119, Server::start 95, api export 94.
+Non-test unwraps: playback.rs:683, 693 (poisoned lock aborts); cli main.rs:208, 480.
+CI: 8f5b141 green on three runners; main red since bdcaedf on i18n test.
+
+### Scores (CQ Maint Scale Perf Arch Phil Sec Test)
+host 7 6 5 6 6 6 4 7↑
+server 7 8 5 6 8 8 4↓ 6 (rescored: pre-auth fd leak)
+api 8 7 5 5 8 4 5 7
+cli 7 8 – – 7 6 6 2
+Features: Logs 7→8↑; ★Native crash reports 6; Self-update 5→5; Remote API 5→4↓; Monitor stills 7→7; ★Saving/autosave 6; Playback/audio 7→7; ★Merge (host half) 7; Windows data folders 6→6; ★Hardware info 7.
+
+### Good
+Crash line designed properly (stderr + dup'd fd, stack formatting, chains, before --version, real child-process test on 3 runners). Merge recheck before adoption + end-to-end test. Token compare constant-time shared by both transports, OS randomness, required on loopback, line/auth caps, auth timeout (token.rs:29-37, json.rs:35-43). Update file names refused if they escape (updates.rs:399-403 tested); saves sync before rename (projects.rs:183-195).
diff --git a/audit/2026-10-15/media-render-export-core.md b/audit/2026-10-15/media-render-export-core.md
new file mode 100644
index 00000000..b1491546
--- /dev/null
+++ b/audit/2026-10-15/media-render-export-core.md
@@ -0,0 +1,66 @@
+## media, render, export, core: reviewer 3, HEAD 11e1bd3 (includes merged PR #361 / 4f755b4)
+Labels: Verified = read at line; Reasoned = followed path; Suspected = says why.
+Commits in part: render, core none; media 447570a; export fc7e6a9, 4f755b4; 14b5998 (host/window) reviewed for export device.
+
+### Re-verified
+- 12Oct#4 Windows FXC compile unbounded, repeated, Cancel deaf: UNCHANGED. gpu.rs:2486-2560 outside DEVICE_WAIT; cancel only top of frame (concat-export/src/lib.rs:1524), progress after (:1687-1690); Windows export gets off_thread()=sibling() (panes/export.rs:518, preview.rs:152); drawing_in empty shader map (gpu.rs:861, :1267-1291) -> every effect recompiled every export.
+- 12Oct#7 GLES 30s wait ~2.15s: UNCHANGED gpu.rs:1311-1318; wgpu 30 (src/Cargo.toml:35); doc gpu.rs:183-191 wrong.
+- 12Oct#9 HEVC/AV1 export opens NVENC first: WORSE. lib.rs:1485 hardware:true; encoders(true) leads hevc_nvenc/av1_nvenc (encode.rs:117-130). New: background proxy reaches it (N3).
+- 12Oct#10 validation error = lost GPU: UNCHANGED gpu.rs:3401-3437; recovered() second device, WARP, "lost 3 times" (lib.rs:1360-1385).
+- 12Oct#11 combine unscoped, refused passes skipped, lent counter never moves: UNCHANGED gpu.rs:3450-3518, :2552-2558, :3044-3059, :839-842, :1291.
+- 12Oct#31 8-bit phone fallback clips HDR: UNCHANGED gpu.rs:174-181, log only :788-795.
+- 12Oct#33 DEVICE_WAIT doc, two read_back docs: UNCHANGED gpu.rs:183-191, 3312-3321.
+- 12Oct#5 export side failed preview layer dropped silently: UNCHANGED lib.rs:2092-2123.
+- 12Oct§7 encoder retry: FIXED (447570a) except Android launch check never exports (scripts/ci/smoke-android.sh).
+- 11Oct#8 uncaptured GPU errors: PARTLY (rest is #11).
+- 11Oct#11 mux scratch names collide: UNCHANGED lib.rs:1107-1109.
+- 11Oct#12 two decode paths disagree past clip end: UNCHANGED, cause located: decode.rs:1072-1079 ends clip at last pts + OUTPUT frame duration; paced path drops layer (lib.rs:1615-1618), pool holds last frame (pool.rs:999-1008). 24/25p on 30/60p timeline to media end loses last output frame to black ~20% (reasoned).
+- 11Oct#13 pick_rate VFR: UNCHANGED probe.rs:174-178, 230-245; no VFR fixture.
+- 11Oct#14 MAX_SIDE doc: UNCHANGED frame.rs:27-31 vs gpu.rs:849-851.
+- 11Oct#21 CPU fallback docs: UNCHANGED concat-export/src/lib.rs:18-21; gpu.rs:676-678, 1295-1296, 1644-1648.
+- 11Oct#22 dead code core: UNCHANGED timeline.rs:314-321, 497-514.
+- core Neg/ceil on i64::MIN: UNCHANGED time.rs:159-161, 251-259.
+Tally 15: 0 fixed, 1 partly, 13 unchanged, 1 worse.
+
+### Commits
+- 447570a MediaCodec "try again" (#286): does it. send retries via patiently (encode.rs:1300-1327, 1408-1428) 10s patience 2ms pause; flush reads to EOF (:1330-1345, 1441-1455); software encoders unaffected. Two pure tests with fakes (:1640-1760). Misses: never run on phone, no CI export on Android; decoder mirror bug (N9) latent because Android never decodes on MediaCodec (N1); encoder-open reason thrown away (N7).
+- 14b5998 own Vulkan device for export (#324): off_thread (preview.rs:142-153) sibling except Vulkan; reasoning right. Also used by consolidation. No test; rule inline. No effect on Android (no window device, concat/src/gpu.rs:15). Untested on Linux NVIDIA, the second-device road #202 found deadly on Windows NVIDIA laptops.
+- fc7e6a9 consolidation: export part sound: effect_time_offset validated (lib.rs:1011-1015), quantise_at keeps grid (resolve.rs:226, 450-455), EffectClock shifts shader clock only (lib.rs:1734-1770); 4 host tests (concat-host/tests/export.rs:2005-2290) + 2 unit. Misses: EffectClock doesn't forward lost()/deliver_hdr() (N8); HDR consolidation on Android fails (HEVC 10-bit request concat-host/src/consolidate.rs:118-127, MediaCodec filtered from 10-bit encode.rs:551-561); one-line message.
+- 4f755b4 centre transitions: correct, tested (lib.rs:790-830, extend :868-884, 6 unit tests replace 5, host wipe test updated). Misses: chained-lane bug kept (N2); sound crossfades in file only (N5); tail from media_duration = container length (longest stream) - 1 frame (:851-862), a phone file whose sound outlasts picture runs past last picture (suspected, Low); doc lib.rs:531-535 not rewrapped.
+
+### New findings
+HIGH
+- N1 Android never decodes on MediaCodec; every phone clip CPU-decoded (reasoned from FFmpeg decoder tables). Device::attach uses decoder::find(context.id()) (hardware.rs:305-306) = built-in software decoder; open_codec opens same (decode.rs:741-743). FFmpeg's built-in h264/hevc hw configs never list MediaCodec (only separate *_mediacodec decoders) -> hardware_format None -> silent software decode. Yet hardware_wins says "MediaCodec takes everything" (hardware.rs:200-202), log "video decode on MediaCodec" (:358). Effect: 4K HEVC / 10-bit HDR phone footage CPU-decoded for playback/scrub/filmstrip/export: stutter, heat, battery; up to 8 warm readers (pool.rs:568-593) with frame threads on every core (decode.rs:731-735) on a 4 GB phone. No test (only macOS device test hardware.rs:534-544). Fix: open h264_mediacodec/hevc_mediacodec by name, buffer mode; fix N9 first.
+- N2 Chain of transitions climbs lanes into next track's lane (verified by reading). overlap reads outgoing lane at resolve (lib.rs:806), incoming one lane above (:816). A->B->C on track t (lane 2t): B 2t+1, C 2t+2 = track t+1's lane for C's whole length; 4th clip 2t+3. Lanes used as-is (resolve.rs:217); clip_on_track_at one clip per lane (timeline.rs:472-480) -> wherever C overlaps a title/caption/overlay on track t+1, one isn't drawn. Monitor same (lib.rs:2194-2195). Slideshow with cross-fades + auto-captions. Predates 4f755b4; no test of two transitions in a row.
+- N3 Importing >1080p HDR clip opens NVENC in background (worsens #9). Auto proxy for >HD (concat/src/studio.rs:3005-3032); HLG/PQ proxy HEVC hardware:hdr (concat-host/src/proxy.rs:224-243) -> hevc_nvenc first (encode.rs:117-130). On #309 hybrid-GPU laptops adding a phone 4K HDR clip segfaults in libcuda with unsaved edits. On Android proxy fails (10-bit filter encode.rs:551-561) so no proxy. Only escape CONCAT_HARDWARE_ENCODE=0, unmentioned.
+MEDIUM
+- N4 Audio mix holds every later clip's decoded sound; Cancel ignored. mix_to_file opens all pieces at once (audio.rs:405-470), round-robin (:638-685); adelay releases nothing until D -> ~whole timeline in filter queues, ~1.4 GB/hour 48k stereo float; OOM on phone. Speed-curved clip up to 400 inputs (lib.rs:1234). cancelled() only before/after (lib.rs:1130-1136), bar parked at 0 "mixing audio". (reasoned)
+- N5 Transitions crossfade sound in file, not in window. Playback audio from session.flattened_clips() (concat/src/studio.rs:2663, 10544-10560; concat-host/src/session.rs:301-307) never through resolve_transitions (private lib.rs:550). Export fades both sides (lib.rs:812-822) and now plays outgoing past out-point. Editor hard cut; file crossfade + unheard source audio. (verified)
+- N6 Android export sheet opens up to three MediaCodec encoders on UI thread. panes/export.rs:631-633 encodable() for H.264/HEVC/AV1 at first publish; each opens a 720p MediaCodec encoder (encode.rs:170-180, 472-500). Slower cold start; vendor codec hang -> ANR. (path verified, cost suspected)
+- N7 FFmpeg's own error lines discarded. Log relay keeps only FATAL/PANIC as errors; ERROR/WARNING -> debug, dropped (ffi.rs:44-56, 101-105). Encoder refusal logs only "{name} would not open ...: Invalid argument" (encode.rs:1163). On #286 the explanation is thrown away.
+LOW
+- N8 EffectClock doesn't forward lost/deliver_hdr (lib.rs:1734-1770; compositor.rs:51-59 defaults). Latent.
+- N9 Decoders drop a packet on EAGAIN (decode.rs:895-897, audio.rs:372-374, samples.rs:276-278). Harmless now; corrupts until keyframe once N1 fixed.
+- N10 ~45 MB (8-bit) / ~80 MB (HDR) allocated per 4K exported frame: read_back (gpu.rs:3352-3380), write_frame (encode.rs:1489-1508) copies twice.
+- N11 Prefetch doc wrong ("single-threaded" prefetch.rs:240-246); every reader frame-threads on all cores (decode.rs:723-735); 4 workers x all cores, 8 warm readers.
+- N12 Android SDR proxy holds a MediaCodec encoder (proxy.rs:223-240); phones allow 1-2 sessions -> export mid-proxy may fail (suspected).
+- N13 Volume doc drift: Animation::volume says factor (animate.rs:243-247; volume_at :280-283 no caller); mixer/playback absolute (audio.rs:61-70; concat-host/src/playback.rs:256-264).
+
+### Numbers
+media 10 328 (+262) tests 102 (+2) unsafe 38 unwrap 210 (23) panic 0
+render 7 723 (0) tests 91 unsafe 1 unwrap 86 (9) panic 0
+export 4 240 (+201) tests 35 (+1) + 15 host export tests (+4) unsafe 0 unwrap 31 (0) panic 0
+core 2 897 (0) tests 49 unsafe 0 unwrap 26 (4) panic 2 (arena.rs:131, 187 unreachable) + 6 asserts time.rs
+Biggest files: gpu.rs 3 561, export lib.rs 3 404 (+194), encode.rs 2 436 (+263), gpu/tests.rs 1 856, decode.rs 1 757, pool.rs 1 676.
+Biggest fns: drawing_in 433 (gpu.rs:861), mix_to_file 303 (audio.rs:392), render_picture 268 (lib.rs:1428), open_video 238 (encode.rs:588), resolve_transitions_at 195 (lib.rs:554), Encoder::open 179.
+All 38 media unsafe carry SAFETY; FFmpeg struct mirrors match 8.1. Shipped app panic=abort ([profile.app]) so any export-thread expect kills window; none looked reachable from user data. UI-thread blocking: encodable() (N6).
+
+### Scores (CQ Maint Scale Perf Arch Phil Sec Test)
+core 8 8 6 7 8 8 7 7: no commits.
+media 7 6 4↓ 6↓ 7↓ 7 7 7: 447570a right; N1, N4, N7.
+render 7 5 6 7 8 7 7 7: no commits; #4,#7,#10,#11 stand.
+export 6↓ 4 5↓ 5 7 7 7 7: two contributor features well tested; lib.rs 3 404; N2; wrapper; one audio input per piece.
+Features: Export desktop 8→7↓; Export on phone 6→6; Hardware encode 6→5↓; Hardware decode 7→5↓; Colour/HDR 7→6↓; Transitions 7→6↓; Crop/transform/blend/masks 7→7; Keyframes 7→7; Proxies 5→4↓; Playback decode 7→6↓.
+
+### Good
+447570a right fix, pure helpers tested with fakes, patience from last packet, stall = stated error. Contributor PRs with real tests (fc7e6a9 end-to-end phase+grid through reopened project; 4f755b4 tests for new rule incl. keys at instants). 14b5998 diagnoses cause, records per-API threading rule with issue numbers. No non-test unwraps in export; FrameRate::checked; sizes checked against device limits. render/core held still; GPU parity suite intact.
diff --git a/audit/2026-10-15/project-effects-ml-phones-ci.md b/audit/2026-10-15/project-effects-ml-phones-ci.md
new file mode 100644
index 00000000..9408877f
--- /dev/null
+++ b/audit/2026-10-15/project-effects-ml-phones-ci.md
@@ -0,0 +1,98 @@
+## project, effects, text, vision, speech, Android, iOS, CI and release: reviewer 4, HEAD ab21457
+Baseline 12 Oct at c06560e. Every commit in this part read as a commit; CI, release and issue state via gh. Nothing edited; cargo not run. Metrics script reproduced the 12 Oct table exactly on c06560e (git archive).
+
+### Re-verified (12 Oct numbering)
+- 1 Release notes' previous tag: FIXED (31ac1d5). release.yml:174-178 excludes v*-* for a final tag; 0.2.7 notes span 0.2.6 to 0.2.7.
+- 2 iPhone Import: UNCHANGED, shipped in 0.2.7. platform.rs:691; only concat-android/src/lib.rs:223 installs a picker; #332; 4 159 .ipa downloads.
+- 3 Phones not built on push: UNCHANGED, worse in effect. mobile.yml:53-77 no push trigger; hand runs 76ee423, 7c5a572 red; PR runs for fc7e6a9, a68641c, 7cb4494 died "workflow file issue" (38805162579, 38814642023, 38735268530). No phone build of main since 14b5998.
+- 4 Windows blur/glow FXC; GPU suites skip on WARP as passed: UNCHANGED; gpu/tests.rs:41-48. 0.2.7 notes say "a blur or a glow no longer stalls an export" (be1b882, reverted).
+- 12 Phone that fails to start vanishes; iPhone plist keys: UNCHANGED. concat-android/src/lib.rs:388-391, 395-397; platform.rs:459-461; src/scripts/ios-app.sh:54-97 lacks UIFileSharingEnabled, LSSupportsOpeningDocumentsInPlace.
+- 13 locales.py --check passes with keys missing: UNCHANGED, proven (scripts/locales.py:193-194; vn.json 11 short, red only on parity).
+- 14 CI internals in release notes: PARTLY. release-notes.sh:46 filters fix(release|ci); :35/:56 types with digits. 0.2.7 notes still carry the self-contradicting DXC pair, "_exit... rc.2's arm64" (fix(linux)), "BUILT_IN holds 16 locales", "Update Japanese translations". Draft not hand-edited.
+- 15 Windows ships CPU ORT, speech asks DirectML: UNCHANGED concat-speech/src/lib.rs:39-41, chatterbox.rs:303-314, concat-vision/src/runtime.rs:145-148.
+- 16 Intel Mac ORT: WORSE, confirmed in field (#333, #342; H4). API 27 vs 23 note (build-app.yml:91-96) still for 13.3+.
+- 17 --smoke stops before window: UNCHANGED concat/src/lib.rs:1996-1998; scripts claim "through the first frame" (smoke-windows.ps1:17-18, smoke-linux.sh:10-11, smoke-macos.sh:10-11). #339, #365 are this class.
+- 18 ISA guard can't catch #193 repeat: UNCHANGED isa-allow.txt:56-63 PROVISIONAL; Ivy Bridge emulator runs --version only (smoke-windows.ps1:177-179).
+- 19 Bundle workflow tag-only: UNCHANGED build-app.yml:45-46 workflow_call.
+- 20 Signing policy vs workflow: worse in reality: 0.2.7 Windows unsigned (H2).
+- 21 Dependabot cargo group red: UNCHANGED dependabot.yml:27-36; #327 holds 18 updates.
+- 28 Concurrency cancels main: UNCHANGED ci.yml:31-32; six runs 16 Oct 05:55-06:01 cancelled.
+- 29 ubuntu-latest -> Ubuntu 26 on 19 Oct: UNCHANGED, 4 days. ci.yml:37, mobile.yml:96, 262, release.yml:42, 118, nix.yml:35, models.yml:41.
+- §7 smoke-ios exit status: FIXED smoke-ios.sh:193. smoke-android trap: FIXED :118. exec bit: FIXED (100755). Simulator env inline: UNCHANGED mobile.yml:485-496; mobile-env.sh:17 only 2 device targets. "Record emulator first result as baseline": WORSE (baseline is "never runs app code", H3).
+- Phone downloads no digest: UNCHANGED ffmpeg-mobile.sh:38-39; sherpa-mobile.sh:38-39, 56-57.
+- Cards skip with no adapter: UNCHANGED concat-host/src/cards.rs:602.
+- project AddClip whole media: UNCHANGED commands/clips.rs:34-38. Speed curve grows over neighbour: UNCHANGED commands/properties.rs:285-305 (SetClipSpeed ripples :108-118, curve doesn't). DOCUMENT_VERSION 1 fine (doc.rs:41; newer refused doc.rs:142, test lib.rs:1691). All tests one file: WORSE lib.rs 4 294 -> 4 533, 121 tests.
+- effects never through D3D12: UNCHANGED; Linux coverage now gone too (H1).
+- android three bridge crash paths, 0 tests: UNCHANGED; emulator never reaches it; new paths M1, H5.
+
+### Commits
+- 31ac1d5 notes from previous release: works. Misses CI-internal subjects with product scope (fix(linux), fix(windows)) and the be1b882/3178b78 contradiction; not hand-edited.
+- 76ee423 phone builds started on emulator/simulator: iOS genuinely good (v0.2.7 "smoke: ok", exit checked, app up 60 s, first log line). Simulator build is separate; shipped .ipa never launched; newest runtime only. Android never ran app code (H3); mobile.yml:33 claim false for Android.
+- 7c5a572 / 0660c01 emulator crash "not tested": diagnosis plausible (ORT cpuinfo_arm_linux_init under ndk_translation); not hiding real-phone crash (untestable_here returns early unless translating, smoke-android.sh:242). But turns gate into permanent no-op reporting success; API 30 blind matcher (:262-266) accepts any SIGSEGV at 0x8 before first line. "Started on a real phone by hand" has no record.
+- b483314 0.2.7 without Windows ARM: honest; #326 open cause unknown; arm64 Windows cell blank, nothing says x64 runs emulated. Release run cancelled and v0.2.7 tag MOVED to 14b5998 (runs 38430320304, 38431729500).
+- e525d7c thin LTO Intel Mac: landed 18:04 after v0.2.7 build (16:35); 0.2.7 Intel is fat LTO; unproven; optimises a bundle that doesn't start on its declared minimum.
+- 447570a encoder retry: patient final drain as asked; never against MediaCodec in CI; #286 unconfirmed.
+- b607491 exit 1: CI-only benefit.
+- c71f900 speech stages: clean (JobStage, transcribe.rs:423-436, tts.rs:884-903); zero-window safe; one test; tts::speak 313 lines.
+- fc7e6a9 ConsolidateClips: best-validated command (clips.rs:893-1004: grid, edge transitions, content underneath, detached audio, full selection); 7 pure tests incl. undo. Two non-test expects (clips.rs:634 safe, :976). Merged with failing PR CI (38805162274); host test consolidation_preserves_the_timeline_clock_of_animated_effects hits 300 s TIMEOUT on WARP (run 38815298005), lacks heavy_gpu guard. Orphans (L4).
+- a68641c: window/locales only; merged on red; touches ui/phone/editor.slint with no phone build.
+- 5add799 (#364): every effect name English in 24 locales; deliberate.
+
+### Release and CI state
+- v0.2.7 published 13 Oct 18:23 from 14b5998 after tag moved; 13 bundles. Downloads: Windows setup 17 077 + msi 1 838; Android apk 16 001; iOS ipa 4 159; macOS arm64 2 701; macOS x86_64 640; Linux ~2 580. First release with every desktop launch check passing (Linux arm64 _exit fixed; Windows x86_64 through emulated Ivy Bridge). Windows UNSIGNED ("No signing certificate: the build is unsigned.", log 17:32:45). macOS ad-hoc, not notarised. Android signed with real key (V3, CN=Concat, SHA-256 923308...6e2e). Android smoke API 30 and 35 "NOT TESTED", exit 0.
+- main red since 14 Oct 05:05 (f064b68, #304 vn.json). 7 failed + 7 cancelled since; last green e525d7c. Engine: locales --check fails at ci.yml:141, before clippy (:150) and tests (:168). Windows: i18n test (concat/src/i18n.rs:447) + fc7e6a9 test timeout. Six PRs merged on red incl. two contributor features.
+- Pipeline weaknesses: tag-only (19), phones not on push (3), Android gate no-op (H3), signing skips silently (H2), no check bundled libraries' minimum OS vs declared (H4).
+- Supply chain: every action SHA-pinned; desktop downloads digest-pinned; phone downloads not; ci.yml, mobile.yml, build-app.yml no permissions: block (L2).
+- 106 open issues (82 on 12 Oct); 6 open PRs (19).
+
+### Open issues that matter
+Windows start-up: #339 closes ~5 s after start (undiagnosed, no log asked); #365 main access violation on hybrid NVIDIA/AMD right after "concat cards" thread starts (suspected second device/driver; CI can't see: WARP + --smoke); #306, #311 Smart App Control (unsigned keeps open); #326 Windows ARM crashes on --version; #313, #318 AVX2 (probably fixed, unconfirmed, open); #297 FXC stalls; #291 DirectML.
+Intel Mac: #333 open, #342 closed by reporter, no fix.
+Android: #312 Android 11 preview (fix shipped, unconfirmed); #286 export codecs (unconfirmed); #353, #328 safe area; #351, #354, #355, #356 layout clipping; #229, #221, #151.
+iOS: #332 Import dead; #338 how to install; #294 crash at launch iOS 16.6; #299 fix shipped, unconfirmed.
+
+### New findings
+HIGH
+- H1 main no Linux tests, clippy or effect-package GPU run for two days; features merged anyway. ci.yml:141 aborts before :150/:168; lavapipe is the only CI place the packages render (Windows skips gpu/tests.rs:41-48, macOS compiles). Eleven commits untested on Linux incl. fc7e6a9, a68641c. Windows also fails (i18n test, consolidation timeout). Fix vn (vi exists); stop merging on red.
+- H2 0.2.7 Windows shipped unsigned; nothing stops it. build-app.yml:946-949 exits 0 without WINDOWS_CERTIFICATE_PFX, on a tag too (verified in log). SAC refuses unsigned DLLs (#306, #311); SmartScreen warns everyone. Fail tag build without cert, or say "unsigned" in notes.
+- H3 Android launch check never tested the app and gates release as passed. Emulators x86_64 (mobile.yml:295-309), APK arm64-only; every Android smoke (rc.3, rc.4, v0.2.7) ended untestable_here exit 0 (smoke-android.sh:241-277). 16 001 APK downloads never started by CI; #312, #286 are what it was added for. Build x86_64 APK from same source for emulator, or device farm.
+- H4 Intel Mac 0.2.7 doesn't start on macOS 12.0-13.2. dyld Symbol not found __ZNSt3__18to_charsEPcS0_d from libonnxruntime.1.23.2.dylib (libc++ symbol from 13.3). Declares 12.0 (build-app.yml:102, LSMinimumSystemVersion :582), links runtime at load (:294-302, ORT_PREFER_DYNAMIC_LINK=1). Launch check macos-15-intel only. 2015-2017 Macs on Monterey crash (#333, #342). On 13.3+ cutouts/Chatterbox face API 27 vs 23. Load ORT lazily (load-dynamic) or raise minimum to 13.3; add otool -l minimum-OS check over every dylib.
+- H5 (suspected, from android-activity 0.6.1 and Slint) Android closes whenever activity recreated. concat-android/Cargo.toml:66 only orientation|keyboardHidden|screenSize. Dark mode (incl. scheduled), split screen, display size, font scale, locale recreate NativeActivity; android-activity calls android_main again (its lib.rs:66-79). Second call: set_var with worker threads alive (lib.rs:31-39, SAFETY false), slint::android::init fails (platform set), return (:388-391) finishes activity. App vanishes mid-edit. Add uiMode|screenLayout|smallestScreenSize|density|fontScale|locale|layoutDirection; guard second android_main.
+MEDIUM
+- M1 (suspected) Android import crashes after process kill or provider error. ConcatFiles Fragment via in-memory dex (lib.rs:154-195), added for pick (ConcatFiles.java:146-150); restore after kill -> ClassNotFoundException, pick (in-memory PENDING) lost. copy() catches only IOException (:262-289); SecurityException from openInputStream (revoked grant, cloud URI) kills app.
+- M2 Android projects/imports wiped on uninstall, unreachable. HOME = external_data_path (lib.rs:37-38); every pick copied to getExternalFilesDir/Imported (ConcatFiles.java:203-204). Android 11+ hides Android/data; "folder the user can reach" false. Reinstall deletes everything; imports double storage.
+- M3 Release notes told ~19 000 Windows users a false thing (14 above). Published, never corrected.
+- M4 Shipped iOS app never launched. Simulator smoke newest runtime only (smoke-ios.sh:131-148); MinimumOSVersion 15.0 (ios-app.sh:21, 67). Nothing checks prebuilt SherpaOnnxC/FFmpeg minimums vs 15.0 (H4 class). #294 open.
+- M5 Nobody knows if main builds for phones today; a68641c changed phone editor without a phone build.
+- M6 (suspected) Cutouts no CPU fallback on accelerator run-time failure (concat-vision/src/runtime.rs:62-80, 115-121), unlike speech's #291 retry. NNAPI (deprecated Android 15) and CoreML never in CI.
+- M7 No offline shader portability check: naga only wgsl-in (concat-effects/Cargo.toml:15); packages reach D3D12 and GLES first on users' machines. Pure naga hlsl-out/glsl-out test over every package, no GPU needed.
+LOW
+- L1 concat.light-leak-flash/effect.wgsl:17 pow(leak_pos - progress, 2.0) negative base, undefined in WGSL/HLSL/MSL; NaN band on a driver that doesn't fold it. Write d*d. (suspected)
+- L2 No top-level permissions: in ci.yml, mobile.yml, build-app.yml.
+- L3 v0.2.7 tag re-pointed after cancelled run, tagged while 14b5998's CI running.
+- L4 Consolidation orphans renders (concat/src/studio/consolidate.rs:101-112); HDR merge 10-bit HEVC hardware-first -> 12 Oct NVENC crash reachable from a right-click.
+- L5 concat-text/src/lib.rs:569 mmaps fonts; replaced/truncated font = SIGBUS. Known.
+- L6 English word-for-word lines 993 -> 1 357 (CI log); effect names now always English (5add799).
+
+### Numbers
+project 10 685 (+400) tests 121 (1) (+7) unsafe 0 unwrap 376 (13) (+2) panic 6
+effects 7 256 (0) ~1 785 WGSL, 123 packages (125 after #361) tests 72 (9) unsafe 0 unwrap 134 (5) panic 2
+text 2 607 tests 24 (1) unsafe 1 unwrap 67 (1)
+vision 2 760 tests 31 (8) unsafe 0 unwrap 22 (2)
+speech 3 411 (+104) tests 22 (3) (+1) unsafe 3 unwrap 23 (2)
+android 398 + 291 Java tests 0 unsafe 10 unwrap 1 (1)
+Biggest fns: commands/clips.rs:12 apply 670; commands/properties.rs:12 apply 338; tts.rs:719 speak 313; concat-text/src/lib.rs:1201 paint 302; catalogue.rs:193 from_manifest 220. project lib.rs 4 533 lines, all 121 tests.
+CI: workflows 2 396 lines (mobile.yml +201, build-app.yml +11, release.yml +8); scripts/ci 1 350 (+576: two phone smokes, Show-Crash smoke-windows.ps1:65-84). Model downloads fail closed on missing/wrong digest (concat-host/src/models.rs:295-310).
+
+### Scores (CQ Maint Scale Perf Arch Phil Sec Test)
+project 8 6 6 6 8 7 6 8
+effects 8 7 6 5 8 8 6 6↓
+text 7 6 5 6 8 8 7 6
+vision 8 8 5 5 9 8 8 7
+speech 7 5 6 6 7 7 8 6
+android 6↓ 7 – 7 6↓ 7 7 1
+CI and release: 4↓ (was 6)
+Features: Effects/filters 8→7↓; Templates 7→7; LUT 8→8; Custom/community packages 7→7; Cutout/brush 6→5↓; Enhance 5→5; Captions engine 7→7; TTS 7→7; Voiceover 6→6; Export on phone 6→6; ★Phone launch checks Android 2, iOS 7; Bundle launch checks 6→6.
+
+### Good
+iOS simulator smoke is real (exit checked, 60 s watch, first log line), passed on v0.2.7. 31ac1d5 works. ConsolidateClips refuses rather than loses, rechecks before adopting (concat-host/src/consolidate.rs:51-68), one undo step, tests prove each refusal. Model downloads fail closed (models.rs:295-300). Effect packages can't reach outside: FFmpeg filter allowlist (filters.rs:22-190), LUT plain .cube beside manifest (manifest.rs:534-537), reserved built-in author. Show-Crash records faulting module from Windows event log (smoke-windows.ps1:65-84). Android release signing correct, verified on shipped APK.
diff --git a/audit/2026-10-15/window.md b/audit/2026-10-15/window.md
new file mode 100644
index 00000000..92b851c7
--- /dev/null
+++ b/audit/2026-10-15/window.md
@@ -0,0 +1,71 @@
+## Window crate (src/crates/concat): reviewer 1, HEAD ab21457
+
+Counting reproduced 12 Oct exactly on c06560e (23 923 lines, 37 (4) unwraps, 13 unsafe, 87 tests). Did not run cargo; ran locales.py --check; read CI via gh.
+
+### Re-verified (12 Oct numbering)
+- 2 iPhone Import does nothing: UNCHANGED. platform.rs:675-694 only log::warn (:691). install_file_picker only at concat-android/src/lib.rs:223; install_export_publisher :321; nothing on iOS.
+- 5 >24 titles drop out mid-drag: UNCHANGED. studio.rs:2733-2749 hold_still per title -> preview.rs:169 -> concat-media/src/pool.rs:541 (STILLS_HELD=24).
+- 6 Shared save scratch: UNCHANGED. concat-host/src/projects.rs:174; saves from studio.rs:2633 and close_project :7300-7301.
+- 8 Captions sheet transcribes a clip it isn't showing: UNCHANGED, one site worse. select.slint:127, segmented-control.slint:43, inspector/controls.slint:559 still self-assign; captions.slint:118,132,143,216,225 as were; 34 one-way bindings into these (new: settings.slint:932).
+- 9 studio.rs growth: WORSE. 11 161 (+295). First split landed: studio/consolidate.rs 213.
+- 12 Phone that can't start vanishes; iPhone files invisible: PARTLY. b607491 logs + exit 1 (lib.rs:2027-2028); report_startup_failure still no-op on phones (platform.rs:445-461); ios-app.sh lacks UIFileSharingEnabled / LSSupportsOpeningDocumentsInPlace.
+- 13 locales.py --check passes with missing keys: UNCHANGED. scripts/locales.py:193-194 never sets failed.
+- 17 --smoke stops before window exists: UNCHANGED lib.rs:1996-1999.
+- 18 Window smalls: UNCHANGED host.rs:360-366, dock.rs:19-20, meters-pane.slint:198,206,214.
+- 22 4px drag threshold < finger slop: UNCHANGED timeline/lanes.slint:591-592.
+- 23 #151 after Browse: UNCHANGED panes/start.rs:165,175,215,243.
+- 24 Detach unmutes deliberately muted video: UNCHANGED studio.rs:10522-10523.
+- 25 Unknown argument refuses to start: UNCHANGED launch.rs:42-47.
+- 26 Commits that close issues: WORSE. e042957 "Fixes #335", 3479407 "Closes #320" closed both; neither in a release (0.2.7 shipped at 14b5998).
+- 27 RTL laid out LTR: UNCHANGED.
+- 33 Doc drift: UNCHANGED studio.rs:2259-2265, :7632, :10335, :1026 (doc for absent label_of atop shelves). concat/src/gpu.rs:62-66, 96-100 .ok()? drops adapter/device errors unlogged.
+- 11th#3 Paste keeps shape: PARTLY FIXED (e042957). Footage copies right (copy_commands studio.rs:1948-2027, tested). Titles/shapes carry only style, duration, offset_y (:6993-7017): scale, x, rotation, opacity, keys, fades, effects, blend lost. cutout, audio_stream, enhanced never copied. close_project (:7286) leaves clipboard full.
+- 11th Literals through I18n: UNCHANGED + new. studio.rs:3695, 6784, 7278-7279, 9679 ("Freeze frame", "F" shortcut skipping platform::keys), :4815 "Title", :1021 "New title", :1047 "Other"; Merge refusal reasons (M5).
+- 12th platform.rs:347-350 SAFETY untrue: UNCHANGED + new L1.
+Tally: 0 fixed, 2 partly, 13 unchanged, 3 worse.
+
+### Commits since 12 Oct
+- b607491 exit 1 when event loop can't run: does it; phone user still sees nothing; no test.
+- 14b5998 own device on Vulkan (#324): off_thread() for export (panes/export.rs:518), cards (studio.rs:7481), merge (studio/consolidate.rs:116). On Vulkan incl. Android (gpu.rs:51-57) every export and every batch of cards opens a fresh wgpu device (cards.rs:404-407). Never run on a phone. No test.
+- 8f5b141 crash handler: window side fine (call_once).
+- c71f900 staged progress: panes/pace.rs pure, 5 tests. Progress redraws every frame while live (L3). No run id on progress messages (L8 suspected). All 25 locales translated by maintainer.
+- 3479407 + 5c698aa interface scale (#320, contributor): 4 pure tests, runtime untested. Scale absolute not relative (M1). False SAFETY on runtime set_var (L1). 3 keys English in 24 locales. Phone gate right. Never built for phones (H2).
+- e042957 copy of split clip (#335): correct, tested at speed 1 and 2. Duplicate still two undo steps (M4). Used "Fixes".
+- fc7e6a9 consolidate selected clips (contributor): staleness recheck before adopting (concat-host/src/consolidate.rs:52-66), cancel, Drop guard cleanup, window + host export test. REMOVES lossless MergeClips path (M5). Progress only in tray label. Refusals English. 10 keys English in all locales. merge_blocked in publish_lanes (L2). Not on phones. No issue ref.
+- a68641c caption style to all titles on track: one undoable batch, tested undo/redo. Silent when nothing changes; strips receiving titles' position keys (L4). English in 23 locales.
+- 5add799 effect names out of locales (#364): does what it says; en.json 673 keys. Edited vn.json without noticing it isn't compiled in.
+- 5a623c0/013d284 contributor vn.json, merged via f064b68 (PR #304): BROKE main (H1).
+- bdcaedf ClearType: sound design (LCD text only on opaque swapchain); never looked at on a Windows screen (L7).
+
+### New findings
+HIGH
+- H1 main red since 14 Oct; Linux clippy+tests not run since. vn.json stale second Vietnamese locale, not in BUILT_IN (i18n.rs:57-83), 11 keys behind. locales.py --check exits 1 on parity (:157, :229). i18n::tests::every_locale_file_ships_and_each_once fails (CI Windows run 38815298005). On Engine job the failed locale step skips clippy, tests, doc tests and gRPC on Linux. Every completed push since f064b68 failed/cancelled; #329, #336, #357, #359 and 5add799 landed on red tree. Fix: delete vn.json or fold into vi.json. (verified by running)
+- H2 No window change since v0.2.7 built for Android or iOS. mobile.yml fails as "workflow file issue" on every contributor PR (3479407, e042957, fc7e6a9, a68641c); last two hand-started runs on main failed (76ee423, 7c5a572). (verified via gh)
+MEDIUM
+- M1 Interface scale replaces display scale instead of multiplying. lib.rs:169-183, 1486-1530; prefs.rs:15-18, 141-148; hint en.json:416. Slint winit treats SLINT_SCALE_FACTOR as absolute (i-slint-backend-winit-1.18.1/winitwindowadapter.rs:602-608, 666-668). Windows laptop at 150%: "100%"/"125%" makes UI smaller; Retina "100%" halves it. While set, moving to another monitor no longer rescales (:1544). Fix: multiply by winit scale_factor, re-apply on scale event. (verified against Slint source)
+- M2 Export sheet can show a codec it won't use. dialogs/export.slint:392-403 (no disabled list), select.slint:125-128 assigns current, panes/export.rs:171-176 refuses codec device can't encode. Pick AV1 without AV1 encoder (most Android, many Windows): picker says AV1, file is H.264. Since 8d3bf19 (28 Sep).
+- M3 Phones: dialogs and toasts ignore safe area. Only main layout inset (app.slint:594-597). Modals at window root full size (app.slint:677-906), surface up to height-24px (modal.slint:48-57). ToastLayer 16px above bottom (toast.slint:86-88). Edge-to-edge Android (targetSdk 35)/iPhone: toast under nav bar/home indicator, sheet header/close under status bar. (code verified; device effect reasoned)
+- M4 Duplicate is two undo steps; first leaves whole source file on timeline. studio.rs:6885-6890 place_copy(..., None); place_copy :6986-7040 applies AddClip (whole file at lane end) then shaping Batch separately (concat-project/src/editor.rs:187-205). Ctrl+Z after duplicating 3s of a 2h file leaves a 2h clip. Fix: pass a gesture like Paste.
+- M5 Merge re-encodes where it used to rejoin losslessly. fc7e6a9 dropped MergeClips path. merge() (studio/consolidate.rs:71-200) renders to media/merged-*/merged.mp4, H.264 crf16 fast or HEVC for HDR (concat-host/src/consolidate.rs:118-131). Refused for audio/image clips, clips above other visible content, edge transitions, off-grid cuts (concat-project/src/commands/clips.rs:894-1000). Progress only in tray label, no tooltip (timeline/tray.slint:277-288). Refusals English (consolidate.rs:49-52 error.to_string(); host :34, :62).
+LOW
+- L1 runtime set_var with false SAFETY: lib.rs:1498-1507 while decoder/audio/scheduler/export threads run; Linux getenv race in FFmpeg/libva/Mesa. "System" removes a hand-set value against :166-167. +2 unsafe (:178, :1499).
+- L2 merge checks every lanes publish: merge_blocked() in publish_lanes (studio.rs:8010-8015) every playback tick/drag move; consolidation_clips clones every selected clip (clips.rs:905-912). selection.iter().any per visible clip (studio.rs:7843) O(visible x selected).
+- L3 progress bar redraws every frame while live (primitives/modal.slint:333-338, 373); on phone competes with whisper. (reasoned)
+- L4 "Apply to all" silent and destructive: no feedback when nothing changes (studio.rs:5000-5017), strips OffsetX/OffsetY keys (:1373-1379), button shows when title alone on lane (inspector/text-panel.slint:172-188).
+- L5 contributor PRs add English to every locale: 3479407 (3), fc7e6a9 (10), a68641c (2). 811 values still English across 24 locales.
+- L6 vn.json edited three times while never compiled in.
+- L7 ClearType unseen on Windows; three SystemParametersInfoW calls per frame.
+- L8 captions cancel/restart collide (suspected): no run id (panes/captions.rs:161-170); late Finished(Err cancelled) closes reopened sheet.
+- L9 merge interrupted by quit leaves media/merged-* (Linux _exit lib.rs:2046-2055 skips Drop guard).
+
+### Numbers
+Rust 24 885 (+962). studio.rs 11 161 (+295, 45%); impl Studio :2093-10497 8 405 lines, 219 methods, 94 fields. start() lib.rs:162-2031 1 870 lines (+98). Next: publish_lanes 389, publish_chrome 280, stage_dragged 262, key_editor_rows 255, settings update 247, menu_bar 224. unwrap/expect 59, 4 non-test (host.rs:171, panes/start.rs:152, studio.rs:2711, :9006); panic 0. unsafe 15 (+2). Tests 96 in 22 files (+9). Slint 30 081 (+180); 34 one-way bindings. Locales 26 files (25 + stray vn), 673 keys, placeholders match, --check exits 1.
+
+### Scores
+concat (Rust): CQ6 Maint4 Scale5 Perf6 Arch7 Phil6 Sec6 Test5
+concat (Slint): CQ7 Maint7 Scale6 Perf6↓ Arch6 Phil6
+locales: CQ6↓ Maint8 Scale7 Arch8 Phil8 Test7
+Features: Launcher 8→8; Import 8→8; Timeline editing 8→7↓; Copy/paste 6→7↑; Undo 8→8; Titles 8→8; Auto-captions 7→7; TTS 7→7; Export 8→7↓; Export on phone 6→6; Phone shell 6→5↓; Languages 8→7↓; Themes 8→8; ★Interface scale 4; ★Merge selected clips 6; ★Caption style across track 7; Windows text (ClearType) 5→5; Command-line launch 7→7; Accessibility 2→2.
+
+### Good
+pace.rs pure with 5 tests; studio/consolidate.rs first feature out of studio.rs, Drop guard, adopt only if inputs unchanged, epoch-tied; e042957 precise with failing-without-fix test; a68641c test covers undo/redo/other lane/keys/missing id; 5add799 locales smaller and honest; the gate caught vn.json, nobody acted; non-test unwrap 4, panic 0 in 25k lines.
diff --git a/audit/AUDIT-2026-10-15-QUICK.md b/audit/AUDIT-2026-10-15-QUICK.md
new file mode 100644
index 00000000..b2c3cc60
--- /dev/null
+++ b/audit/AUDIT-2026-10-15-QUICK.md
@@ -0,0 +1,100 @@
+# Concat audit, 15 October 2026: the quick version
+
+The one-screen scorecard for [`AUDIT-2026-10-15.md`](AUDIT-2026-10-15.md).
+Code at `ab21457` (export at `11e1bd3`), 36 commits past the 12 October
+audit, with v0.2.7 released in between. The reviewers' own reports are in
+[`2026-10-15/`](2026-10-15/).
+
+## In one paragraph
+
+0.2.7 is the first release through every desktop launch check, and about
+45 000 people downloaded it in two days. The new code is mostly careful and
+tested. But `main` has been red for two days over one stray file, so Linux
+has run no tests since 14 October and six PRs merged on top. Android, the
+second-largest audience, has never been started by CI, decodes everything
+on the CPU, and probably closes on a dark-mode switch. iOS ships without
+Import. Windows shipped unsigned. The Intel Mac build does not start on
+macOS 12. Phones have not been built since the release.
+
+## Scores
+
+| Aspect | 12 Oct | Today |
+|---|:-:|:-:|
+| Code quality | 6 | 6 |
+| Design architecture | 8 | 7 ↓ |
+| Design philosophy | 6 | 5 ↓ |
+| Maintainability | 5 | 5 |
+| User likeability | 7 | 6 ↓ |
+| Performance | 6 | 5 ↓ |
+| Scalability | 5 | 5 |
+| Security | 6 | 5 ↓ |
+| Testing | 6 | 5 ↓ |
+| Documentation | 5 | 6 ↑ |
+| CI and release | 6 | 4 ↓ |
+
+## The numbers
+
+| | Today | Δ |
+|---|---:|---:|
+| Rust lines (`src/` only) | 93 423 | +3 061 |
+| Tests | 778 | +30 |
+| `unsafe` | 110 | +40 (crash handler 32, machine info 6, scale 2) |
+| Non-test `unwrap`/`expect` | 68 | +2 |
+| `studio.rs` | 11 161 lines | +295 |
+| `start()` | 1 870 lines | +98 |
+| `cargo test --workspace` | 773 passed, 1 failed (`vn.json`), 1 ignored | |
+| fmt, clippy (here) | pass | |
+| `locales.py --check` | exit 1 (`vn.json`) | |
+| CI on `main` | red since 14 Oct 05:05; last green `e525d7c` | |
+| Open issues / PRs | 106 / 6 | +24 / −13 |
+
+## Carried findings
+
+1 fixed (release notes' previous tag) plus the four uncommitted-work asks,
+about 5 partly, about 7 worse, the rest unchanged. None of the 11 October
+server items has landed.
+
+## New, High
+
+| # | What | Where |
+|---|---|---|
+| H1 | `main` red two days; Linux ran no tests, clippy or effect GPU render; six PRs merged on red | `locales/vn.json`, `ci.yml:141` |
+| H2 | Windows 0.2.7 unsigned; signing step exits 0 without a certificate | `build-app.yml:946-949` |
+| H3 | iOS ships without Import; Files app can't see the folder either | `platform.rs:675-694`, `ios-app.sh` |
+| H4 | Android launch check never runs the app, reports success | `mobile.yml:301`, `smoke-android.sh:241-277` |
+| H5 | Android probably closes on dark mode, split screen, font or locale change (suspected) | `concat-android/Cargo.toml:66`, `lib.rs:388-391` |
+| H6 | Intel Mac 0.2.7 does not start on macOS 12.0-13.2 (#333) | `build-app.yml:102, 294-302` |
+| N1 | Android never decodes on MediaCodec; every clip CPU-decoded | `hardware.rs:305-306`, `decode.rs:741-743` |
+| N2 | A chain of transitions climbs into the next track and hides it | `concat-export/src/lib.rs:806, 816` |
+| N3 | Importing a large HDR clip opens NVENC in the background | `proxy.rs:224-243`, `encode.rs:117-130` |
+| N4 | No phone built since v0.2.7 | `mobile.yml:53-77` |
+
+## New, Medium
+
+M1 interface scale replaces the display's scale (smaller UI on most
+laptops) · M2 export sheet shows a codec it won't use · M3 phone dialogs
+and toasts under the system bars · M4 Duplicate is two undo steps and
+leaves the whole file · M5 Merge re-encodes instead of a lossless join ·
+M6 Remote API fd exhaustion without a token · M7 audio mix holds the
+timeline's sound, Cancel ignored · M8 transition sound crossfades only in
+the file · M9 FFmpeg's error lines dropped · M10 Android import crash after
+process kill · M11 Android projects deleted on uninstall · M12 encoder
+probes on the UI thread at start · M13 crash handler takes locks · M14 a
+second GPU device per export on Vulkan and Android · M15 0.2.7 notes claim
+a reverted fix · M16 README promises gRPC and a CLI no download ships.
+
+## Do first
+
+1. Delete `vn.json`; guard the consolidation test; stop merging on red.
+2. Pin `ubuntu-24.04` before 19 October.
+3. Android `config_changes` and a guarded second `android_main`.
+4. Fail a release without a Windows certificate; get one.
+5. A real Android launch check (x86_64 APK or device farm).
+6. Interface scale multiplies the display's scale, before it ships.
+7. Intel Mac: lazy ONNX Runtime or a 13.3 minimum, plus an `otool` check.
+8. Android `*_mediacodec` decoders.
+9. iOS document picker and plist keys; README iOS 🚧 until then.
+10. Phone workflow on PRs and `main`.
+11. Transition lanes from the original track, with a three-clip test.
+12. No NVENC first for proxies.
+13. Correct the 0.2.7 notes; drop gRPC and the CLI from the README.
diff --git a/audit/AUDIT-2026-10-15.md b/audit/AUDIT-2026-10-15.md
new file mode 100644
index 00000000..a965c5f9
--- /dev/null
+++ b/audit/AUDIT-2026-10-15.md
@@ -0,0 +1,540 @@
+# Concat audit, 15 October 2026
+
+Where the codebase stands three days after the 12 October audit. Code read
+at `ab21457`, with the two crates PR #361 touched (concat-export, the host's
+export tests) read at its merge `11e1bd3`. That is 36 non-merge commits past
+the last audit's `c06560e`, with v0.2.7 released in between (13 October,
+from `14b5998`). Same method, same 1-10 scales, every carried finding
+re-verified against today's lines. Every finding names a file and a line. The sibling
+`AUDIT-2026-10-15-QUICK.md` is this audit as a one-screen scorecard, and
+`2026-10-15/` holds the four reviewers' own reports.
+
+**How it was done.** Four reviewers each read one part of the tree: the
+window (Rust, Slint, locales); host, server, API and CLI; media, render,
+export and core; project, effects, text, vision, speech, Android, iOS, CI
+and release. Each re-verified the 12 October findings in its part, read
+every commit since as a commit (does it do what it says, with what test,
+missing what), scored its part and reported new findings. Each reviewer's
+counting script reproduced the 12 October table exactly on `c06560e` before
+counting today's. The headline items (main's CI, the Android decoder, the
+Android launch check, the unsigned Windows release, the Intel Mac crash,
+the interface scale, the transition lanes) were then re-read by hand, and
+run where a command could show it. `cargo fmt --check`, `cargo clippy
+--workspace --all-targets -D warnings`, `cargo test --workspace`,
+`scripts/locales.py --check` and `scripts/models.py --check` were run on an
+Apple silicon Mac (macOS 26, Metal). The app was not launched, and no phone
+or release bundle was built.
+
+**Labels.** *Verified*: read at the line, or run. *Reasoned*: followed
+through the code or a dependency's documented behaviour, not reproduced.
+*Suspected*: says why.
+
+---
+
+## The short version
+
+0.2.7 is the first release in which every desktop bundle passed its launch
+check, and about 45 000 people downloaded it in two days, 19 000 of them on
+Windows and 16 000 on Android. The code that landed since is mostly careful
+and mostly tested. None of that is the problem.
+
+The problem is that the project has stopped looking at what it ships:
+
+- **`main` has been red for two days** over one stray file (`vn.json`), and
+ because that check runs first, Linux has run no clippy, no tests and no
+ effect-package GPU render since 14 October. Six PRs, two of them
+ contributor features, were merged on top of red.
+- **Android, the second-largest audience, has never been started by CI.**
+ The launch check runs an arm64-only APK on x86_64 emulators, gets
+ "untestable", and reports success. On the device, video is decoded on the
+ CPU while the log says MediaCodec, and the activity probably closes
+ itself on a dark-mode switch.
+- **iOS ships without Import.** 4 159 people downloaded an editor that
+ cannot open a single clip.
+- **Windows 0.2.7 is unsigned.** The signing step exits 0 when the
+ certificate is missing, so Smart App Control refuses the app (#306, #311)
+ on the platform with the most users.
+- **The Intel Mac build does not start on macOS 12 to 13.2**, which it
+ declares as supported (#333).
+- **Phones have not been built since v0.2.7.** The phone workflow fails on
+ every contributor PR with "workflow file issue", and both hand-started
+ runs on `main` failed.
+
+Of roughly 80 findings carried from the last two audits, one was fixed
+outright (the release notes' previous tag), about five partly, about seven
+got worse, and the rest are unchanged. None of the 11 October server items
+has landed.
+
+Most of the fixes are small. Deleting `vn.json` turns `main` green.
+Failing the tag build without a certificate is three lines. The Android
+recreation fix is one manifest attribute. The order is in section 8.
+
+---
+
+## 1. The numbers
+
+### 1.1 Size, tests and hygiene per crate
+
+Counted the 12 October way: Rust lines are `src/` only, without `build.rs`
+and `tests/` directories; tests include `tests/`; brackets are counts
+outside test modules. Deltas are against 12 October.
+
+| Crate | Rust lines | Δ | Slint / WGSL / Java | Tests (files) | `unsafe` | `unwrap`/`expect` (non-test) | `panic!` family (non-test) |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| concat (window) | 24 885 | +962 | 30 081 slint (+180) | 96 (22) | 15 | 59 (4) | 0 |
+| concat-host | 11 971 | +1 129 | | 99 (22) | 42 | 286 (2) | 0 |
+| concat-project | 10 685 | +400 | | 121 (1) | 0 | 376 (13) | 6 |
+| concat-media | 10 328 | +262 | | 102 (11) | 38 | 210 (23) | 0 |
+| concat-render | 7 723 | 0 | shaders inline | 91 (5) | 1 | 86 (9) | 0 |
+| concat-effects | 7 256 | 0 | ~1 785 wgsl, 125 packages | 72 (9) | 0 | 134 (5) | 2 |
+| concat-export | 4 240 | +201 | | 35 (3) | 0 | 31 (0) | 0 |
+| concat-speech | 3 411 | +104 | | 22 (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 | 517 | +3 | | 3 (1) | 0 | 2 (2) | 0 |
+| concat-android | 398 | 0 | 291 java | 0 (0) | 10 | 1 (1) | 0 |
+| **Total** | **93 423** | **+3 061** | | **778** | **110** | **(68)** | **10** |
+
+Where it went: the host's crash handler and machine header (+1 129, of
+which `crash.rs` is 556 lines and 32 of the 38 new `unsafe`), the window
+(+962: interface scale, the copy fix, merge, caption style, the staged
+progress model), project's `ConsolidateClips` (+400), media's MediaCodec
+retry (+262) and export's consolidation clock and centred transitions
+(+201). 30 tests landed on 3 061 lines.
+
+- `studio.rs` is **11 161 lines** (+295, 45 % of the window's Rust).
+ `impl Studio` is 8 405 lines, 219 methods over 94 fields. `start()` in
+ `lib.rs:162-2031` is **1 870 lines** (+98). The first feature kept out
+ of it landed: `studio/consolidate.rs`, 213 lines.
+- `concat-export/src/lib.rs` is 3 404 lines (+194); `encode.rs` 2 436
+ (+263); `gpu.rs` 3 561. concat-project's 121 tests are all in one
+ 4 533-line `lib.rs`; its `commands/clips.rs` `apply` is 670 lines.
+- Every `unsafe` carries a SAFETY comment. Three are not true as written:
+ `platform.rs:347-350` (carried), `concat/src/lib.rs:1498` (runtime
+ `set_var` while decoder, audio and export threads run) and
+ `concat-host/src/crash.rs:213-214` (`dladdr` takes the loader's lock).
+- The shipped profile is `panic = "abort"`, so every non-test `expect` and
+ every lock `unwrap` on a worker thread ends the whole app.
+
+Other facts: 125 effect packages (57 filters, 21 effects, 22 transitions,
+25 audio). 25 shipped locales at 673 keys after #364, plus the stray
+`vn.json`; placeholders match English in every file. 1 357 values are the
+English word for word (the CI log's count, which now includes effect names
+by design). 106 open issues (82 on 12 October) and 6 open PRs (19).
+
+### 1.2 The checks, run today
+
+| Check | Result |
+|---|---|
+| `cargo fmt --all --check` | pass |
+| `cargo clippy --workspace --all-targets -D warnings` | pass (here; CI's Linux clippy has not run since 14 Oct) |
+| `cargo test --workspace` at `9a528c5` | **773 passed, 1 failed, 1 ignored** in 30 suites. The failure is `i18n::tests::every_locale_file_ships_and_each_once` (`vn.json`) |
+| `scripts/locales.py --check` | **exit 1**: `vn.json` is not in `BUILT_IN` and has no README |
+| `scripts/models.py --check` | pass: 23 models, 3.37 GB, mirrored |
+| CI on `main` | **red since `f064b68` (14 Oct 05:05)**. Last green: `e525d7c` (13 Oct 18:04). The engine job fails at `ci.yml:141` (locales), before clippy (`:150`) and the tests (`:168`). The Windows job fails the same i18n test, and fc7e6a9's consolidation test hits its 300 s timeout on WARP |
+| Phone builds | Not built since v0.2.7. `mobile.yml` has no push trigger; PR runs die with "workflow file issue" (38830606380 for #361, also for fc7e6a9, a68641c); the last two `workflow_dispatch` runs on `main` (76ee423, 7c5a572) failed |
+| Release v0.2.7 | 13 bundles built; every desktop launch check passed. Windows **unsigned** (log: "No signing certificate: the build is unsigned."); macOS ad-hoc, not notarised; Android signed with the real key; Android smoke "NOT TESTED", exit 0 |
+
+---
+
+## 2. Scorecard for the whole app
+
+| Aspect | 12 Oct | Today | One line |
+|---|:-:|:-:|---|
+| Code quality | 6 | 6 | New code is mostly pure helpers with tests, and non-test unwraps stay at 68 in 93k lines. Against that: two new SAFETY comments that are false, a crash handler whose doc oversells it, commits that still say "Fixes", and `studio.rs` +295. |
+| Design architecture | 8 | 7 ↓ | The interface scale overrides the display's scale through an environment variable set at run time. Merge replaced a lossless join with a re-encode. On Vulkan, which includes Android, every export and card round opens a second GPU device. Chained transitions climb lanes. |
+| Design philosophy | 6 | 5 ↓ | Gates that report success without testing (the Android smoke, the WARP-skipped GPU suites), merges on a red `main`, a signing step that skips quietly on a release tag, and release notes that told Windows users a stall was fixed when its fix had been reverted. |
+| Maintainability | 5 | 5 | `studio.rs` 11 161, `start()` 1 870, export `lib.rs` 3 404, project's tests in one 4 533-line file. One feature split out of `studio.rs`. |
+| User likeability | 7 | 6 ↓ | Fades, centred transitions, staged progress with time left and the copy fix are real gains. But iOS cannot import, the Intel Mac app does not start on macOS 12, Windows is unsigned, and "Interface scale" makes the interface smaller on most laptops. |
+| Performance | 6 | 5 ↓ | Android decodes everything on the CPU (N1). The audio mix holds the timeline's decoded sound in memory. A second Vulkan device per export on phones. The progress bar repaints the window every frame while it runs. |
+| Scalability | 5 | 5 | Unchanged: the fd per connection, the one hub thread, the per-sample mixer scan, the per-edit bin clone. |
+| Security | 6 | 5 ↓ | An unsigned Windows release. Any local process can exhaust the window's file descriptors through the Remote API without a token. The update chain trusts a digest from the same manifest as the URL and strips quarantine. No `permissions:` block in three workflows. |
+| Testing | 6 | 5 ↓ | The crash handler's child-process test on three runners and consolidation's end-to-end tests are good. But Linux has run no tests for two days, the Android gate never runs the app, the shipped iOS app is never launched, and the effect packages have never compiled through D3D12 or GLES. |
+| Documentation | 5 | 6 ↑ | Today's pass took MCP out of the READMEs, fixed the effect count, corrected THIRD_PARTY_NOTICES against what CI ships, and removed the em dashes. Still wrong: the README promises a gRPC API and a CLI that no download ships, `concat-api/src/lib.rs:7` names MCP, four docs promise a CPU fallback. |
+| CI and release | 6 | 4 ↓ | 0.2.7 is the first release through every desktop launch check, with notes that start at the right tag. But `main` is red with Linux skipped, phones are not built, the Android gate is a no-op, Windows shipped unsigned, the Intel Mac shipped unable to start on its declared minimum, and the tag was re-pointed after a cancelled run. |
+
+---
+
+## 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 12 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 still wrong, dead `Project`/`source_duration`, `i64::MIN` edges. |
+| concat-project | 8 | 6 | 6 | 6 | 8 | 7 | 6 | 8 | `ConsolidateClips` is the best-validated command in the crate, with 7 tests. The speed curve still grows over a neighbour; one 4 533-line test file; a 670-line `apply`. |
+| concat-media | 7 | 6 | 4 ↓ | 6 ↓ | 7 ↓ | 7 | 7 | 7 | The MediaCodec encoder retry is right and tested with fakes. But the Android decode path cannot reach MediaCodec (N1), the mix holds the timeline's sound (M7), and FFmpeg's own error lines are thrown away (M9). |
+| concat-render | 7 | 5 | 6 | 7 | 8 | 7 | 7 | 7 | No commits. The FXC compile, the GLES wait, validation-as-loss and `combine`'s missing scope all stand. |
+| concat-effects | 8 | 7 | 6 | 5 | 8 | 8 | 6 | 6 ↓ | Filter allowlist, LUT path check and reserved author are solid. Off Linux CI since 14 Oct; never compiled through D3D12 or GLES; no offline portability check. |
+| concat-export | 6 ↓ | 4 | 5 ↓ | 5 | 7 | 7 | 7 | 7 | Two contributor features, both well tested. `lib.rs` 3 404 lines; chained transitions climb lanes (N2); a wrapper that does not forward its methods. |
+| concat-host | 7 | 6 | 5 | 6 | 6 | 6 | 4 | 7 ↑ | A real crash test on three OSes and end-to-end merge tests. The crash handler takes locks its doc says it does not; the save scratch and held stills as they were; the update chain unsigned. |
+| concat (Rust) | 6 | 4 | 5 | 6 | 7 | 6 | 6 | 5 | New code mostly pure and tested; a first split from `studio.rs`. Against: a false SAFETY claim, the scale (M1), Duplicate (M4), Merge in English, and nothing proven on phones. |
+| concat (Slint) | 7 | 7 | 6 | 6 ↓ | 6 | 6 | – | – | A continuous repaint while jobs run; 34 one-way bindings into self-assigning primitives, one of them live in the export sheet (M2); modals and toasts outside the safe area (M3). |
+| locales and `locales.py` | 6 ↓ | 8 | 7 | – | 8 | 8 | – | 7 | #364 is principled and the files are honest. But a stray `vn.json` has kept CI red for two days, missing keys still pass, and contributor PRs copy English into every locale. |
+| concat-api | 8 | 7 | 5 | 5 | 8 | 4 | 5 | 7 | No code commits. A raw `ConsolidateClips` is now reachable with no verb that renders a selection; captions are window-only; `export.run` overwrites. |
+| concat-server | 7 | 8 | 5 | 6 | 8 | 8 | 4 ↓ | 6 | No commits; rescored, not regressed: an unauthenticated connection leaks an fd, which ends in failed saves and a spinning accept loop. No auth throttle. |
+| concat-cli | 7 | 8 | – | – | 7 | 6 | 6 | 2 | Gained the crash handler; still no logger, no signal handling, `render` writes in place. |
+| concat-speech | 7 | 5 | 6 | 6 | 7 | 7 | 8 | 6 | Clean stages and time left. DirectML still "asked for" on a Windows build that ships the CPU runtime. |
+| concat-vision | 8 | 8 | 5 | 5 | 9 | 8 | 8 | 7 | No commits. No CPU retry when an accelerator fails at run time; NNAPI and CoreML never run in CI. |
+| concat-text | 7 | 6 | 5 | 6 | 8 | 8 | 7 | 6 | No commits. |
+| concat-android | 6 ↓ | 7 | – | 7 | 6 ↓ | 7 | 7 | 1 | Activity recreation unhandled (H5); the dex-loaded fragment cannot survive a process restore and the copy thread can crash (M10); projects deleted on uninstall (M11); no test reaches this code. |
+
+---
+
+## 4. Scores per feature
+
+| Feature | 12 Oct | Today | Why |
+|---|:-:|:-:|---|
+| Launcher, new project, recents | 8 | 8 | No commits; #151 returns after Browse. |
+| Import, bin, drag and drop | 8 | 8 | Desktop and Android fine. iPhone Import still does nothing (H3). |
+| Timeline editing | 8 | 7 ↓ | Merge now re-encodes (M5); Duplicate leaves the whole file behind one undo (M4); the 4 px drag threshold is under a finger's slop. |
+| Copy and paste | 6 | 7 ↑ | A copy of a split clip shows its own frames (#335, tested). Titles and shapes still lose scale, position, rotation, keys and effects; cutout and audio stream are never copied. |
+| Undo and redo | 8 | 8 | The caption-style batch is one step, tested. Duplicate is two. |
+| Transitions | 7 | 6 ↓ | Centred on the cut and tested (#361); Fade to Black and Fade to White added. But a chain of transitions climbs into the next track (N2), sound crossfades in the file and not in the window (M8), and `combine` is unscoped. |
+| Effects and filters | 8 | 7 ↓ | Names are English in every language (#364, intended). D3D12 and GLES unproven; Linux CI off since 14 Oct; the 0.2.7 notes claim the Windows blur stall is fixed. |
+| Colour and HDR | 7 | 6 ↓ | HDR proxies fail on Android and open NVENC first on desktop (N3); the 8-bit phone fallback still clips. |
+| Colour grading | 7 | 7 | Unchanged. |
+| Titles and text | 8 | 8 | "Apply caption style to all" added; it is silent and strips position keys (L4). |
+| Keyframes | 7 | 7 | Extending a clip keeps its keys at their instants (#361). Keys still slide on a trim. |
+| Crop, transform, blend, masks | 7 | 7 | Unchanged. |
+| Cutout, smart brush | 6 | 5 ↓ | The Intel Mac app does not start on 12-13.2 (H6); DirectML is pretend; no CPU retry; never run on a phone in CI. |
+| Enhance | 5 | 5 | Same runtime as cutout. |
+| Auto-captions | 7 | 7 | Staged bar with time left. The sheet can still transcribe a clip it is not showing. |
+| Text-to-speech | 7 | 7 | Staged bar; DirectML semantics unchanged. |
+| Voiceover | 6 | 6 | Untouched. |
+| Export (desktop) | 8 | 7 ↓ | The codec picker can disagree with the file (M2); chained transitions hide a track (N2); the audio mix cannot be cancelled and grows with the timeline (M7). |
+| Export on a phone | 6 | 6 | The encoder retry shipped, but no phone export has ever run in CI. Encoder probes block the UI thread at start (M12). #286 unconfirmed. |
+| Hardware encode | 6 | 5 ↓ | NVENC is now reached by background proxies, not only by an export (N3). |
+| Hardware decode | 7 | 5 ↓ | Windows and Apple fine; Android never uses MediaCodec (N1). |
+| Proxies | 5 | 4 ↓ | HDR proxies open NVENC first and fail on Android; Android SDR proxies may hold the phone's only encoder session. |
+| Playback and audio | 7 | 7 | Unchanged: per-sample clip scan, frees on the real-time thread, a mono device loses the right channel. |
+| Monitor stills | 7 | 7 | More than 24 live titles still drop out mid-drag. |
+| Remote API | 5 | 4 ↓ | Pre-auth fd exhaustion (M6); weak tokens and no throttle; `stop` freezes the UI; README promises gRPC and a CLI nobody can download. |
+| Self-update | 5 | 5 | Unchanged chain: no signature, no Authenticode check, quarantine stripped, downgrades allowed. |
+| Logs | 7 | 8 ↑ | Machine, CPU and commit at the top of every log; a native crash line proven in CI. |
+| ★ Native crash reports | – | 6 | The right shape. Takes the loader's lock and can allocate (M13); two codes in its table can never arrive. |
+| ★ Interface scale | – | 4 | Absolute, not relative (M1); a hint that says the opposite; a `set_var` race; desktop only. |
+| ★ Merge selected clips | – | 6 | Stale-checked, cancellable, cleaned up, tested end to end. Replaced the lossless join, shows progress only in a tray label, refuses in English, and orphans its renders. |
+| ★ Caption style across a track | – | 7 | One undo step, tested; silent, and strips position keys. |
+| Phone shell | 6 | 5 ↓ | iOS Import dead; modals and toasts under the system bars (M3); activity recreation closes the app (H5); not built since v0.2.7. |
+| ★ Phone launch checks | – | Android 2, iOS 7 | iOS: a real simulator launch, but not the shipped .ipa or the minimum OS. Android: "not tested" every time, reported as passed. |
+| Bundle launch checks (desktop) | 6 | 6 | First release with every desktop bundle through. Missed #333 (minimum OS), #339 and #365 (Windows start-up). |
+| Languages | 8 | 7 ↓ | 25 languages, honest files. CI red over `vn.json`; English from contributors; no right-to-left layout. |
+| Accessibility | 2 | 2 | Unchanged. |
+
+---
+
+## 5. The 12 October findings, re-verified
+
+| Status | Count | Which |
+|---|---:|---|
+| Fixed | 1 + 4 asks | #1 release notes' previous tag (31ac1d5). The uncommitted-work asks: the encoder retry with a patient final drain (447570a), `smoke-ios` checks exit status, `smoke-android` has a trap, smoke scripts have the exec bit. |
+| Partly | ~5 | #12 a phone that cannot start now exits 1 (but still vanishes for the person); #14 CI internals in notes (filter added, 0.2.7 notes still carry four); 11 Oct #3 paste (footage right, titles and shapes not); 11 Oct #8 GPU errors; 11 Oct #21 MCP docs (READMEs fixed, `concat-api/src/lib.rs:7` not). |
+| Worse | ~7 | #9 NVENC first (now reached by background proxies); #16 Intel Mac runtime (confirmed in the field, #333 #342); #26 commits closing issues (e042957 "Fixes #335", 3479407 "Closes #320", neither released); #3 phones not built on push (now not built at all); #8 captions bindings (+1 site); `studio.rs` growth; project's single test file. |
+| Unchanged | the rest | Among them: #2 iPhone Import, #4 Windows FXC compile, #5 24 held stills, #6 shared save scratch, #7 GLES 2.15 s wait, #10 validation read as device loss, #11 `combine` unscoped, #13 missing keys pass, #15 DirectML, #17 `--smoke` stops before the window, #18 instruction-set guard, #19 bundle workflow by tag only, #21 Dependabot red, #22 4 px slop, #23-#25, #27 right-to-left, #28 concurrency cancels `main`, #29 `ubuntu-latest`, and every server item from 11 October (stop cancels first, `Builder::spawn`, write timeout, `exists`/`is_sane`). |
+
+Two of the unchanged ones now have a deadline or a field report:
+
+- **#29**: `ubuntu-latest` becomes Ubuntu 26 on **19 October**, four days
+ from now (`ci.yml:37`, `mobile.yml:96, 262`, `release.yml:42, 118`,
+ `nix.yml:35`, `models.yml:41`).
+- **#2**: 0.2.7 shipped with it; #332 is the field report.
+
+---
+
+## 6. New findings, by severity
+
+### High
+
+**H1. `main` has been red for two days, Linux has run nothing, and features were merged on top.** *Verified by running and in CI.*
+- `src/crates/concat/locales/vn.json` is a second, stale Vietnamese file
+ (from PR #304, merged as `f064b68`). It is not in `BUILT_IN`
+ (`i18n.rs:57-83`), which ships `vi`, and it is 11 keys behind.
+- `locales.py --check` exits 1 on it at `ci.yml:141`, which runs before
+ clippy (`:150`) and the tests (`:168`). Linux is CI's only place where
+ the 125 effect packages render (lavapipe); Windows skips them and macOS
+ only compiles. So no effect-package GPU run, Linux test or clippy has
+ happened since 14 October.
+- On Windows the same test fails (`i18n.rs:344`), and fc7e6a9's
+ `consolidation_preserves_the_timeline_clock_of_animated_effects` hits
+ the 300 s timeout on WARP because it lacks the `heavy_gpu` guard.
+- Merged on red since: #329, #336, #357 (fc7e6a9), #359, a68641c, 5add799,
+ #361. Three of them edited `vn.json` without noticing it is never
+ compiled in.
+- **Fix:** delete `vn.json`; guard the consolidation test; make "green
+ before merge" a rule. Five minutes.
+
+**H2. Windows 0.2.7 shipped unsigned, and nothing stops the next one.** *Verified.*
+- `build-app.yml:946-949`: when `WINDOWS_CERTIFICATE_PFX` is unset the step
+ prints "No signing certificate: the build is unsigned." and exits 0, on
+ a release tag too. The v0.2.7 log has that line.
+- Smart App Control refuses unsigned DLLs (#306, #311, both open), and
+ SmartScreen warns everyone else, on the platform with 19 000 of the
+ downloads.
+- **Fix:** fail a tag build without the certificate, or say "unsigned" in
+ the notes until there is one.
+
+**H3. iOS ships without Import.** *Verified.*
+- `platform.rs:675-694` only logs "no file picker on this platform yet";
+ the only `install_file_picker` call is Android's
+ (`concat-android/src/lib.rs:223`). Nothing on iOS installs a picker or
+ an export publisher.
+- `src/scripts/ios-app.sh` still lacks `UIFileSharingEnabled` and
+ `LSSupportsOpeningDocumentsInPlace`, so the Files app cannot see the
+ app's folder either. There is no way to get a clip in.
+- 4 159 downloads of the 0.2.7 `.ipa`; #332 confirms it. The README marks
+ iOS ✅.
+
+**H4. The Android launch check has never run the app, and it gates releases as passed.** *Verified.*
+- The emulators are x86_64 (`mobile.yml:301`), the APK is arm64-only, and
+ ARM translation crashes ONNX Runtime's static initialiser. Every Android
+ smoke (rc.3, rc.4, v0.2.7) ended in `untestable_here`
+ (`smoke-android.sh:241-277`) with exit 0. The API 30 matcher accepts any
+ SIGSEGV at 0x8 before the app's first line (`:262-266`).
+- `mobile.yml:33` says "No phone build passes unless it starts on a
+ phone". For Android that is false. The "started on a real phone by hand"
+ step has no record anywhere.
+- So the 16 001 Android downloads, #312 (Android 11) and #286 (export) were
+ never checked by CI.
+- **Fix:** build an x86_64 APK from the same source for the emulator, or
+ use a device farm.
+
+**H5. Android probably closes itself whenever the activity is recreated.** *Suspected, reasoned from android-activity 0.6.1 and Slint.*
+- `concat-android/Cargo.toml:66` declares only
+ `orientation|keyboardHidden|screenSize`. A dark-mode switch (including a
+ scheduled one at sunset), split screen, a display-size, font-scale or
+ locale change recreates the NativeActivity, and android-activity calls
+ `android_main` again in the same process.
+- On that second call `set_var` runs while worker threads are alive
+ (`lib.rs:31-39`), `slint::android::init` fails because the platform is
+ already set, and `return` (`:388-391`) finishes the activity.
+- The app vanishes mid-edit with no message.
+- **Fix:** add `uiMode|screenLayout|smallestScreenSize|density|fontScale|locale|layoutDirection`
+ to `config_changes`, and guard a second `android_main`.
+
+**H6. Intel Mac 0.2.7 does not start on macOS 12.0 to 13.2.** *Verified in the field (#333, #342).*
+- dyld: `Symbol not found __ZNSt3__18to_charsEPcS0_d` from
+ `libonnxruntime.1.23.2.dylib`. That libc++ symbol exists from macOS 13.3.
+- The bundle declares 12.0 (`build-app.yml:102`, `LSMinimumSystemVersion`
+ at `:582`) and links the runtime at load (`:294-302`,
+ `ORT_PREFER_DYNAMIC_LINK=1`). The launch check runs on `macos-15-intel`
+ only, so it cannot see this.
+- **Fix:** load ONNX Runtime lazily (`ort` `load-dynamic`) or raise the
+ minimum to 13.3; add an `otool -l` minimum-OS check over every bundled
+ dylib.
+
+**N1. Android never decodes on MediaCodec; every phone clip is decoded on the CPU.** *Verified by reading; the FFmpeg behaviour reasoned from its decoder tables.*
+- `Device::attach` looks up the hardware config on
+ `decoder::find(context.id())` (`concat-media/src/hardware.rs:305-306`),
+ FFmpeg's built-in decoder, and `open_codec` opens that same decoder
+ (`decode.rs:741-743`). FFmpeg's built-in `h264` and `hevc` decoders list
+ DXVA, D3D11, NVDEC, VAAPI, VDPAU, VideoToolbox and Vulkan, never
+ MediaCodec, which exists only as the separate `*_mediacodec` decoders.
+ Only the encoder side ever names those (`encode.rs:108-110`).
+- Meanwhile `hardware_wins` says "MediaCodec takes everything"
+ (`hardware.rs:198-202`), and the log says "video decode on MediaCodec"
+ (`:358`).
+- A phone's own 4K HEVC and 10-bit HDR footage is decoded on the CPU for
+ playback, scrubbing, filmstrips and export: stutter, heat and battery,
+ with up to 8 warm readers (`pool.rs:568-593`), each with frame threads on
+ every core (`decode.rs:731-735`), on 4 GB phones.
+- **Fix:** open `h264_mediacodec` / `hevc_mediacodec` by name, and first
+ fix N9's dropped packet on "try again", which those decoders will hit.
+
+**N2. A chain of transitions climbs into the next track's lane.** *Verified by reading.*
+- `overlap` reads the outgoing clip's lane at the moment it resolves and
+ puts the incoming clip one lane above (`concat-export/src/lib.rs:806,
+ 816`). With A→B→C on track t (lane 2t), B goes to 2t+1 and C to 2t+2,
+ which is track t+1's lane, for C's whole length. A fourth clip goes to
+ 2t+3.
+- One clip per lane is drawn (`concat-core/src/timeline.rs:472-480`), so
+ wherever C overlaps a title, caption or overlay on the track above, one
+ of them is not drawn. The monitor resolves the same way
+ (`lib.rs:2194-2195`), so the person sees it and cannot fix it.
+- The typical case is a slideshow with cross-fades and auto-captions. It
+ predates #361; #361's fades make chains more likely. No test covers two
+ transitions in a row.
+
+**N3. Importing a large HDR clip opens NVENC in the background.** *Verified.* This makes 12 October's #9 worse.
+- Every video larger than HD gets a proxy (`concat/src/studio.rs:3005-3032`).
+ For HLG or PQ sources it is HEVC with `hardware: hdr`
+ (`concat-host/src/proxy.rs:224-243`), so on Windows and Linux it opens
+ `hevc_nvenc` first (`encode.rs:117-130`).
+- On the hybrid-GPU laptops of #309, adding a phone's 4K HDR clip to a
+ project can segfault the app in libcuda, with unsaved edits, without
+ any export. On Android the proxy fails outright (ten-bit filtering
+ removes MediaCodec, `encode.rs:551-561`), so those clips never get one.
+- The only escape, `CONCAT_HARDWARE_ENCODE=0`, is not mentioned anywhere
+ a user would see it. HDR merges take the same path (L4).
+
+**N4. No phone has been built since v0.2.7.** *Verified via `gh`.*
+- `mobile.yml:53-77` has no push trigger. On pull requests it dies with
+ "workflow file issue" and runs no job (fc7e6a9, a68641c, 7cb4494, #361).
+ The last two hand-started runs on `main` failed.
+- So the scale handler's `cfg`s, the phone "Apply to all" sheet, the
+ progress repaint, the merge code and #361 have never been compiled for
+ the platforms with the most broken experience.
+
+### Medium
+
+**M1. "Interface scale" replaces the display's scale instead of multiplying it.** *Verified against Slint 1.18.1's source.*
+- `prefs.rs:147-149` turns the percentage into `percent / 100`, and
+ `lib.rs:169-183, 1496-1507` sets it as `SLINT_SCALE_FACTOR`, which
+ Slint's winit backend takes as the absolute factor
+ (`winitwindowadapter.rs:602-608`) and which suppresses the monitor's own
+ scale-change event while set (`:1544`).
+- On a Windows laptop at 150 %, "100 %" and "125 %" make the interface
+ smaller. On a Retina Mac, "100 %" halves it. The hint (`en.json:416`)
+ says "larger than the display's own scale". Moving the window to another
+ monitor no longer rescales.
+- **Fix:** multiply by the window's scale factor and re-apply on the
+ scale-change event. Before the next release.
+
+**M2. The export sheet can show a codec it will not use.** *Verified.* `dialogs/export.slint:392-403` passes no `disabled` list, `select.slint:125-128` assigns its own value, and `panes/export.rs:171-176` refuses a codec the device cannot encode. Pick AV1 without an AV1 encoder (most Android phones, many Windows machines): the picker reads AV1, the file is H.264. Present since 28 September.
+
+**M3. On phones, dialogs and toasts ignore the safe area.** *Code verified; on-device effect reasoned.* Only the main layout is inset (`app.slint:594-597`). Modals sit at the window root (`app.slint:677-906`, `modal.slint:48-57`) and toasts 16 px above the bottom edge (`toast.slint:86-88`). On edge-to-edge Android (targetSdk 35) and iPhone, failure toasts sit under the navigation bar or home indicator, and a tall sheet's close button under the status bar. #328, #353.
+
+**M4. Duplicate takes two undo steps, and the first leaves the whole source file on the timeline.** *Verified.* `duplicate` (`studio.rs:6885-6890`) calls `place_copy(..., None)`, which applies `AddClip` (the whole file at the lane's end) and the shaping batch as separate steps (`:6986-7040`). Ctrl+Z after duplicating 3 seconds of a 2-hour file leaves a 2-hour clip at the end of the lane. Pass a gesture, as Paste does.
+
+**M5. Merge now re-encodes where it used to rejoin losslessly.** *Verified.* fc7e6a9 dropped the window's `MergeClips` path. `merge()` (`studio/consolidate.rs:71-200`) renders the selection to `media/merged-*/merged.mp4` at CRF 16 (`concat-host/src/consolidate.rs:118-131`). Rejoining two halves of a split clip costs a render and a generation of quality. Progress shows only in the tray button's label; refusals are English in every language (`consolidate.rs:49-52`). Keep the lossless join when it applies.
+
+**M6. Any local process can exhaust the window's file descriptors through the Remote API, without a token.** *Verified.* Every connection, authenticated or not, leaves a duplicated socket in `connections` until `stop` (`concat-server/src/json.rs:109-122, 164-177`), and the seat is taken before the handshake (`:102, :157`). A few hundred connections (256 on macOS, 1 024 on Linux) and saves fail with "Too many open files" (`concat-host/src/projects.rs:186-187`), decoders fail, and the accept loop spins at 100 % (`:100, :156`). A script that connects once per call gets there in normal use. Off by default.
+
+**M7. The audio mix holds every later clip's decoded sound, and Cancel is ignored while it runs.** *Reasoned.* `mix_to_file` opens every piece at once (`concat-media/src/audio.rs:405-470`) and feeds them round-robin (`:638-685`); a clip's `adelay` releases nothing until the mix reaches its start, so decoded audio piles up: about 1.4 GB per hour of back-to-back clips at 48 kHz stereo float. On a phone that is an out-of-memory kill. A speed-curved clip adds up to 400 inputs (`concat-export/src/lib.rs:1234`). Cancel is checked only before and after (`:1130-1136`).
+
+**M8. Transitions crossfade sound in the file, not in the window.** *Verified.* Playback audio comes from `flattened_clips()` (`studio.rs:2663`, `concat-host/src/session.rs:301-307`), which never passes through `resolve_transitions`. The export fades both sides and, since #361, plays the outgoing clip's sound past its out-point. The editor plays a hard cut; the file has a crossfade and source audio nobody heard while editing.
+
+**M9. FFmpeg's own error lines are thrown away.** *Verified.* The log relay keeps only FATAL and PANIC (`concat-media/src/ffi.rs:44-56, 101-105`). When MediaCodec, NVENC or x264 refuses to open, the log gets "{name} would not open ...: Invalid argument" (`encode.rs:1163`). On the platform where export has never been seen working (#286), the line that explains why is dropped.
+
+**M10. Android import can crash after a process kill or on a provider error.** *Suspected.* `ConcatFiles` is a Fragment loaded from an in-memory dex (`concat-android/src/lib.rs:154-195`) and added for the pick (`ConcatFiles.java:146-150`). If Android kills the process while Photos or Drive is in front (common on 3-4 GB phones), restoring the activity cannot find the class: a crash on return, and the pick is lost. `copy()` catches only `IOException` (`:262-289`), so a `SecurityException` from a revoked grant kills the app.
+
+**M11. Android projects and imports are deleted on uninstall, and users cannot reach them.** *Verified.* `HOME` is `external_data_path` (`lib.rs:37-38`) and every pick is copied into `getExternalFilesDir/Imported` (`ConcatFiles.java:203-204`). On Android 11 and later `Android/data` is hidden from file managers. A reinstall deletes every project, and every import doubles its storage.
+
+**M12. The Android export sheet opens up to three MediaCodec encoders on the UI thread at start.** *Path verified, cost suspected.* `panes/export.rs:631-633` calls `encodable()` for H.264, HEVC and AV1 at the first publish; each opens a real 720p encoder (`encode.rs:170-180, 472-500`). A slow vendor codec delays start; one that hangs heads for an ANR.
+
+**M13. The crash handler takes locks and can allocate, despite its doc.** *Reasoned from the loaders' documented locking.* `dladdr` (`concat-host/src/crash.rs:217`) takes the loader's lock (glibc, bionic and dyld alike); `to_string_lossy` (`:218`) can allocate; on Windows `std::thread::current()` (`:398`) allocates on threads Rust did not start and aborts after TLS teardown, and `GetModuleFileNameW` (`:442`) takes the loader lock. A crash while another thread loads a library, which is exactly #326's start-up window, can hang with nothing logged. Write the raw code and PC first; record module ranges at install.
+
+**M14. On Vulkan, which includes Android, every export, merge and card round opens a second GPU device.** *Path verified, impact suspected.* 14b5998 (titled "linux") makes `Monitor::off_thread` (`concat-host/src/preview.rs:146-153`) return none on Vulkan, and Android picks Vulkan (`concat/src/gpu.rs:52-57`). The cards build a new compositor every round (`cards.rs:404-407`). On 3-4 GB phones and the #312-class drivers, that is a second device's memory and pipeline compiles. It is also the second-device road #202 found deadly on Windows NVIDIA laptops, now untested on Linux NVIDIA.
+
+**M15. Release notes told Windows users a false thing.** The 0.2.7 notes say "a blur or a glow no longer stalls an export" (be1b882), which was reverted. Published, never corrected. Four more CI-internal lines went out with them.
+
+**M16. The README promises a gRPC API and a CLI that no download ships.** *Verified.* Releases build only `-p concat` (`build-app.yml:523`) without the `grpc` feature (`concat/Cargo.toml:133-134`), and no `concat-cli` is published. The window serves JSON-RPC over TCP only.
+
+### Low
+
+- **L1.** Runtime `set_var` with a false SAFETY claim (`concat/src/lib.rs:1498-1507`) while decoder, audio and export threads run; on Linux, FFmpeg, libva and Mesa call `getenv` on those threads. Choosing "System" also removes a hand-set value.
+- **L2.** Merge's checks run on every lanes publish, every playback tick and drag move (`studio.rs:8010-8015`), cloning every selected clip (`clips.rs:905-912`); `selected` per visible clip is O(visible × selected) (`studio.rs:7843`).
+- **L3.** The progress bar repaints the whole window every frame while a job runs (`primitives/modal.slint:333-338, 373`), competing with whisper for CPU and battery on a phone.
+- **L4.** "Apply caption style to all" is silent when nothing changes (`studio.rs:5000-5017`), strips the receiving titles' position keys (`:1373-1379`), and shows when the title is alone on its lane.
+- **L5.** Merged renders are never collected: undo keeps `media/merged-*`, and Clear cache spares `media/`. An HDR merge goes through the NVENC-first path.
+- **L6.** Captions: no run id on progress or finish (`panes/captions.rs:161-170`), so a cancelled run's late "cancelled" can close the reopened sheet (suspected).
+- **L7.** ClearType (`bdcaedf`) ships LCD text on the most-used platform without anyone having looked at it on a Windows screen.
+- **L8.** `EffectClock` does not forward `lost()` or `deliver_hdr()` (`concat-export/src/lib.rs:1734-1770`). Latent.
+- **L9.** Decoders drop a packet on "try again" (`decode.rs:895-897`, `audio.rs:372-374`, `samples.rs:276-278`). Harmless now; corrupts pictures as soon as N1 is fixed.
+- **L10.** About 45 MB (8-bit) or 80 MB (HDR) allocated per exported 4K frame (`gpu.rs:3352-3380`, `encode.rs:1489-1508`).
+- **L11.** API `ConsolidateClips` adopts any file whose duration covers the span (`commands/clips.rs:612-645`), with no verb that renders a selection.
+- **L12.** Weak Remote API tokens accepted (`settings.rs:575`), no auth throttle (`json.rs:267-291`), `0.0.0.0` binds allowed. On a LAN, a short token can be brute-forced into writes anywhere under home.
+- **L13.** Two crash codes in the table can never arrive: fail-fast skips the filter (`crash.rs:358-359`).
+- **L14.** A one-channel output device loses the right channel (`playback.rs:1172-1175`); the audio callback frees memory on the real-time thread (`:1107`).
+- **L15.** Model and update downloads have no size limit (`concat-host/src/models.rs:258-276`).
+- **L16.** `concat.light-leak-flash/effect.wgsl:17` calls `pow` with a negative base, undefined in WGSL, HLSL and MSL (suspected NaN band on some drivers). Write `d*d`.
+- **L17.** No top-level `permissions:` in `ci.yml`, `mobile.yml` or `build-app.yml`.
+- **L18.** The `v0.2.7` tag was re-pointed after a cancelled run, and tagged while its commit's CI was still running.
+- **L19.** No offline shader portability check: naga is used only with `wgsl-in`; `hlsl-out` and `glsl-out` over every package would prove D3D12 and GLES on every runner without a GPU.
+- **L20.** Contributor PRs copy English into every locale (3479407: 3 keys, fc7e6a9: 10, a68641c: 2), and missing keys still pass `--check` (`scripts/locales.py:193-194`).
+
+---
+
+## 7. What is good, and should stay that way
+
+- **The crash line is designed properly**: stack formatting, a duplicated
+ descriptor so it never touches the logger's lock, chained to the
+ previous handler so tombstones and Windows Error Reporting are
+ unchanged, installed before `--version`, and tested by crashing a child
+ process on all three desktop runners.
+- **`ConsolidateClips` refuses rather than loses**: it checks the frame
+ grid, edge transitions, content underneath, detached audio and the whole
+ selection; re-checks inputs before adopting the render; is one undo
+ step; and is tested end to end through a reopened project. The best-tested
+ work of the period, from a contributor.
+- **447570a is the right fix built properly**: `patiently` and `until_ended`
+ are pure and tested with fakes, and a stall becomes a stated error, not
+ a short file.
+- **#361 is careful**: half on each side of the cut, keys kept at their
+ instants, a held frame when there is no footage, and every replaced test
+ rewritten for the new rule.
+- **The iOS simulator smoke is real**: exit status checked, the app watched
+ for 60 s, its first log line required.
+- **Every action is pinned by SHA; desktop downloads and models are
+ digest-pinned and fail closed.**
+- **The locales are honest**: 673 keys, every placeholder matching, effect
+ names out by design.
+- **The core held still**: no churn in concat-core or concat-render, and
+ the GPU parity suite against the CPU reference is intact.
+- **Non-test unwraps: 68 in 93k lines.** 4 of them in the window's 25k.
+
+---
+
+## 8. Where we stand, and what to do in what order
+
+The engine is in better shape than the product around it. The fixes that
+matter most this week are small, and most are in CI and packaging, not in
+the engine.
+
+1. **Make `main` green.** Delete `vn.json` (`vi.json` ships). Add the
+ `heavy_gpu` guard to the consolidation test. Then stop merging on red,
+ contributors included. *Minutes.* (H1)
+2. **Pin the runners before 19 October.** Replace `ubuntu-latest` with
+ `ubuntu-24.04` in the seven places listed in section 5. *Minutes.*
+3. **Stop the Android activity from closing.** Add the missing
+ `config_changes` and guard a second `android_main`. *An hour.* (H5)
+4. **Fail a release without a Windows certificate**, and get one. Until
+ then, say "unsigned" in the notes. (H2)
+5. **Make the Android launch check real**: an x86_64 APK for the emulator,
+ or a device farm. Until then, say "not tested" in the release, not
+ green. (H4)
+6. **Fix the interface scale** to multiply the display's scale, before it
+ ships in a release. (M1)
+7. **Intel Mac**: load ONNX Runtime lazily or raise the minimum to 13.3,
+ and add the `otool` minimum-OS check. Tell #333's reporters. (H6)
+8. **Android hardware decode**: open the `*_mediacodec` decoders by name,
+ after fixing the dropped packet on "try again". (N1, L9)
+9. **iOS Import**: a document picker and the two plist keys. Until then,
+ mark iOS 🚧 in the README, not ✅. (H3)
+10. **Fix the phone workflow** so it runs on PRs and on `main`. (N4)
+11. **Transitions**: assign overlap lanes from the original track, not
+ the already-moved one, with a test for three clips in a row. (N2)
+12. **Hardware encode default**: no NVENC first for proxies, or detect
+ hybrid GPUs; surface `CONCAT_HARDWARE_ENCODE`. (N3)
+13. **Correct the 0.2.7 notes** about the blur stall, and drop gRPC and the
+ CLI from the README until a download ships them. (M15, M16)
+
+After that, the medium list in order of who it hurts: the safe area (M3),
+the codec picker (M2), Duplicate's undo (M4), the lossless join (M5), the
+Remote API fd leak (M6), the audio mix (M7).
+
+---
+
+## 9. Caveats
+
+- The app was not launched, and no phone or release bundle was built here.
+ Everything about phones is from code, CI logs and issue reports.
+- "Suspected" findings (H5, M10, M12, M14, L6, L16) are reasoned from
+ code and dependency behaviour, not reproduced. H5 is the one most
+ worth confirming first, on a real phone, by switching dark mode.
+- N1 is verified in Concat's code; the claim that FFmpeg's built-in
+ decoders never offer MediaCodec is from FFmpeg's decoder tables, not
+ from a run on a phone. A log line from a phone during playback would
+ settle it.
+- Counts follow the 12 October method, which each reviewer reproduced
+ exactly on `c06560e`. Line numbers are at `ab21457`, except
+ concat-export and `concat-host/tests/export.rs`, which are at `11e1bd3`.
+- Since this audit's HEAD, today's documentation pass landed (`2e4c4d5`),
+ #361 was merged (`11e1bd3`), LICENSE-EXCEPTIONS was dropped (`9a528c5`)
+ and `docs/merging-clips.md` was deleted (`792b925`). None of those
+ change a finding above except that the README no longer names MCP.