Skip to content

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

updates

Jareer22a17e94e815
diff --git a/audit/AUDIT-2026-09-28-QUICK.md b/audit/AUDIT-2026-09-28-QUICK.md
index c1ca853e..27063aa8 100644
--- a/audit/AUDIT-2026-09-28-QUICK.md
+++ b/audit/AUDIT-2026-09-28-QUICK.md
@@ -115,7 +115,7 @@ running when this was written; score carried from 23 Sep and adjusted only by ha
|---|---|
| fmt | ✅ 0 files would change (8 on 23 Sep) |
| clippy | ✅ 0 warnings |
-| tests | ⚠️ **647 passed, 2 failed**, 1 ignored (650 in all, 544 on 23 Sep). Both failures are window tests that saw German where they expected English (`format::tests::phrases_are_coarse`: "gestern" for "yesterday"; `panes::speech::tests::a_voice_name_reads_as_a_person_would_say_it`): the process-wide locale leaking between tests in one binary, not a product bug. CI on `main` is green. |
+| tests | ⚠️ **647 passed, 2 failed**, 1 ignored (650 in all, 544 on 23 Sep). Both failures are window tests that saw German where they expected English (`format::tests::phrases_are_coarse`: "gestern" for "yesterday"; `panes::speech::tests::a_voice_name_reads_as_a_person_would_say_it`): the process-wide locale leaking between tests in one binary (confirmed: both pass alone, and the suite passes with `--test-threads=1`), not a product bug. CI on `main` is green. |
| perf --check | ✅ 22 of 22 within budget: GPU composite 2.3 ms (3.2), 720p export 204 fps (115), HDR export 111 fps, 4K HLG deep-for-GPU 140 fps vs 29 tone-mapped on the CPU, 4K HEVC 10-bit hardware 112 fps and software 43, three 4K streams 61 fps, 4K Gaussian blur 6.1 ms at radius 10 and 9.3 at 50 |
| CI on `main` | ✅ green at head; 3 of today's 12 runs were red between 10:12 and 11:02 UTC, fixed by `d1d3265` |
diff --git a/audit/AUDIT-2026-09-28.md b/audit/AUDIT-2026-09-28.md
index b6ab1d80..99057d7e 100644
--- a/audit/AUDIT-2026-09-28.md
+++ b/audit/AUDIT-2026-09-28.md
@@ -23,8 +23,8 @@ export/host and API/server/CLI reviews are in and folded in below; the
media, render/effects, window, Slint, models and CI/docs reviews were still
running when this was written, so those crates' scores are marked
*provisional* (carried from 23 September, adjusted only where a hand check
-justified it), and the test and perf results are marked where they were
-still pending. Each of those sections is replaced as its review lands.
+justified it). Each of those sections is replaced as its review lands. The
+check suite, the perf table and the CI history are complete.
**Scale.** 0 to 10. 8 and up means "keep doing this"; 5 to 7 means "works,
with known debt"; 4 and under means "will bite at the next order of
@@ -90,10 +90,59 @@ them tests), `concat-render/src/gpu.rs` 3 599, `concat-export/src/lib.rs`
|---|---|
| `cargo fmt --all -- --check` | 0 files would change (8 on the 23rd) |
| `cargo clippy --workspace --all-targets` | 0 warnings |
-| `cargo test --workspace --no-fail-fast` | *running when this was written; see the QUICK file's Today's checks row, updated when it finished* |
-| `cargo run -p concat-perf --release -- --check` | *pending the test run* |
+| `cargo test --workspace --no-fail-fast` | 647 passed, 2 failed, 1 ignored (650 in all; 544 on the 23rd) |
+| `cargo run -p concat-perf --release -- --check` | 22 of 22 scenarios within budget |
| CI on `main` | green at head (`b78ee56`, CI and Nix both pass) |
+**Two window tests fail locally; CI is green.** Both failures are tests
+that translate a string and got German where they expected English:
+`format::tests::phrases_are_coarse` sees "gestern" for "yesterday"
+(`concat/src/format.rs:502`) and
+`panes::speech::tests::a_voice_name_reads_as_a_person_would_say_it` finds no
+"American" in a voice's detail (`panes/speech.rs:672`). Since `fc09a61` the
+locale is process-wide and the locales test in the same binary loads every
+shipped file in turn, `de` first; the two tests read whichever locale the
+other thread left behind. Confirmed: both pass when run alone, and the whole
+window suite passes with `--test-threads=1`. A test-isolation bug, not a
+product one. The end-to-end export suite (11 tests, 21 s), the parity suite
+(85 tests, 39 s) and every engine crate pass.
+
+**The perf table**, release build, every scenario inside its budget. New
+rows since the 23rd are marked ★.
+
+| Scenario | Measured | Budget |
+|---|---:|---:|
+| plan one frame of 200 clips on 4 tracks | 0.96 µs | ≤ 200 µs |
+| 200 clip additions and 200 undos | 0.78 ms | ≤ 150 ms |
+| ripple delete of 200 clips out of 400 | 0.20 ms | ≤ 20 ms |
+| save and open a 200-clip document | 1.98 ms | ≤ 60 ms |
+| choose the visible clips of 200 for one publish | 0.21 µs | ≤ 50 µs |
+| read an hour's waveform across 2048 columns | 4.41 µs | ≤ 2000 µs |
+| decode 1080p h264 in software | 1 626 fps | ≥ 120 |
+| decode 1080p h264 on the hardware (VideoToolbox) | 316 fps | ≥ 120 |
+| ★ decode 4K h264, the app's choice (software) | 151 fps | ≥ 30 |
+| ★ decode 4K HEVC 10-bit, the app's choice (VideoToolbox) | 112 fps | ≥ 30 |
+| ★ decode 4K HEVC 10-bit in software | 43 fps | ≥ 24 |
+| ★ decode 3 streams of 4K HEVC 10-bit at once | 61 fps | ≥ 24 |
+| ★ decode 4K HLG tone-mapped on the CPU, 960×540 | 29 fps | ≥ 15 |
+| ★ decode 4K HLG deep for the GPU, 960×540 | 140 fps | ≥ 30 |
+| scrub forward over 90 frames, cold | 0.72 ms/frame | ≤ 25 |
+| scrub back over covered ground | 0.25 µs/frame | ≤ 200 |
+| decodes when an effect knob is turned | 0 | ≤ 0 |
+| compose 1080p, 3 layers, a look, on the GPU | 2.31 ms | ≤ 25 |
+| ★ Gaussian blur over 4K, radius 10, on the GPU | 6.11 ms | ≤ 33 |
+| ★ Gaussian blur over 4K, radius 50, on the GPU | 9.34 ms | ≤ 33 |
+| export a 4 s 720p cut of two overlapping clips | 204 fps | ≥ 30 |
+| ★ export the same cut as HDR: HLG, HEVC 10-bit | 111 fps | ≥ 30 |
+
+Three readings worth keeping: the 720p export went from 115 to 204 fps
+with the CPU composite gone; the deep HDR decode that hands the GPU the
+conversion runs at 140 fps where the CPU tone map ran at 29, the "five
+times faster" of the CHANGELOG; and hardware decode of 1080p h264 is still
+five times slower than software (316 against 1 626) because the frame is
+downloaded and converted on the CPU after the chip decodes it, the
+zero-copy item of ARCHITECTURE §10.
+
**`main` is green.** The last 12 runs on `main` are 7 successes, 3
failures and 2 cancellations. The three failures were this morning's
commits between 10:12 and 11:02 UTC (the roadmap move, the export-device
@@ -541,9 +590,7 @@ Scores are judgements over code, calibrated against the 23 September scores
but not measured against users or another editor. Rows marked *provisional*
in §3 and the three findings marked *pending* in §5 rest on hand checks of
the specific lines named, not on a full read; they are replaced as those
-reviews land. The test and perf results were still running when this was
-written and are reported in the QUICK file's checks table as soon as they
-finished. Findings marked (verified) were read end to end by the reviewer
+reviews land. Findings marked (verified) were read end to end by the reviewer
and the headline ones re-read by hand; the rest rest on one reviewer's
reading and a line number you can open. The app was not launched: "user
likeability" is judged from what the UI code does, not from users.