ci(deps): bump actions/checkout from 4 to 6

.ai/research/2026-05-17/CHANGESET_SUMMARY.md

@@ -0,0 +1,53 @@
+# Changeset Summary - 2026-05-17
+
+## Files Created
+
+| File | Purpose |
+|---|---|
+| `PROJECT_CONTEXT.md` | Canonical committed project memory for future sessions. |
+| `.ai/research/2026-05-17/STATE_OF_REPO.md` | Local reconnaissance memo. |
+| `.ai/research/2026-05-17/MEMORY_CONSOLIDATION.md` | Instruction/memory/doc reconciliation. |
+| `.ai/research/2026-05-17/SOURCE_REGISTER.md` | Local and external source index. |
+| `.ai/research/2026-05-17/RESEARCH_LOG.md` | Search strategy, query classes, saturation notes, and limitations. |
+| `.ai/research/2026-05-17/COMPETITOR_MATRIX.md` | Commercial and open-source competitor comparison. |
+| `.ai/research/2026-05-17/FEATURE_BACKLOG.md` | Raw harvested opportunity backlog. |
+| `.ai/research/2026-05-17/PRIORITIZATION_MATRIX.md` | Scored and tiered roadmap candidates. |
+| `.ai/research/2026-05-17/SECURITY_AND_DEPENDENCY_REVIEW.md` | Dependency, native, release, model integrity, and privacy review. |
+| `.ai/research/2026-05-17/DATASET_MODEL_INTEGRATION_REVIEW.md` | Model, dataset, integration, packaging, and evaluation review. |
+| `.ai/research/2026-05-17/CHANGESET_SUMMARY.md` | Summary of this research artifact changeset. |
+
+## Files Modified
+
+| File | Purpose |
+|---|---|
+| `ROADMAP.md` | Updated last refresh to 2026-05-17 and added Round 7 deep-research synthesis plus new Now-priority rows. |
+| `app/src/main/java/com/novacut/editor/engine/CaptionTranslationEngine.kt` | Fixed the supported-language switch for the new Bergamot model variant. |
+| `app/src/main/java/com/novacut/editor/model/Project.kt` | Added optional clip display name storage needed by compound-clip breadcrumbs. |
+| `app/src/main/java/com/novacut/editor/engine/KeyframeBezierGraph.kt` | Fixed cubic-bezier easing evaluation to solve x(time) before reading y(value). |
+| `app/src/main/java/com/novacut/editor/engine/SpeakerSwitchPlanner.kt` | Fixed initial speaker-angle selection when explicit angle assignments reserve the default angle. |
+| `app/src/test/java/com/novacut/editor/engine/CompoundNavStackTest.kt` | Switched JVM test fixtures from `Uri.EMPTY` to the repo's `FakeUri`. |
+| `app/src/test/java/com/novacut/editor/engine/AdjustmentLayerEngineTest.kt` | Aligned the no-layer plan expectation with the engine's documented whole-clip segment behavior. |
+
+## Local-Only Configuration
+
+`local.properties` was corrected locally so Gradle can find the Android SDK in this workspace:
+
+```properties
+sdk.dir=C:\\Users\\--\\AppData\\Local\\Android\\Sdk
+```
+
+This file is ignored and intentionally not committed.
+
+## Verification
+
+- `git diff --check`: no whitespace errors; Git reported the normal Windows line-ending warning that `ROADMAP.md` LF will be replaced by CRLF the next time Git touches it.
+- `.\gradlew.bat :app:compileDebugKotlin --dry-run`: successful after correcting ignored local `local.properties`.
+- `.\gradlew.bat :app:compileDebugKotlin --no-daemon`: successful.
+- `.\gradlew.bat :app:testDebugUnitTest --no-daemon`: compiles and executes 389 tests, with 388 passing and 1 remaining failure:
+  - `AutoSaveStateTest > deserialize_capsPathologicalRecoveredCollections`
+  - Failure: `java.util.NoSuchElementException at AutoSaveStateTest.kt:287`
+  - Notes: the remaining failure occurs after source compile/test-compile succeeds. Earlier compile blockers and unrelated assertion clusters found during verification were repaired in this changeset.
+
+## Commit
+
+This file is part of the final local commit for the 2026-05-17 research pass.

.ai/research/2026-05-17/COMPETITOR_MATRIX.md

@@ -0,0 +1,57 @@
+# Competitor Matrix - 2026-05-17
+
+## Summary
+
+NovaCut's strongest differentiation is not having the longest checklist. The credible position is: Android-native, local-first, transparent model delivery, pro export/interchange, and privacy-preserving AI tools. Commercial competitors dominate template velocity and cloud/social distribution. Open-source desktop competitors dominate mature NLE workflows. Open-source mobile competitors are thinner, which gives NovaCut room if it keeps execution focused.
+
+## Direct and Adjacent Open-Source Projects
+
+| Project | Type | Activity Evidence | Notable Capabilities | NovaCut Lesson |
+|---|---|---|---|---|
+| OpenCut | Open-source CapCut-positioned editor | GitHub metadata fetched 2026-05-17: 50,904 stars, pushed 2026-05-17, MIT | Modern open-source creator editor positioning, local-first appeal, broad attention | Watch for UX expectations and plugin/template direction, but do not copy architecture blindly because NovaCut is Android-native. |
+| devhyper/open-video-editor | Android open-source editor | GitHub metadata fetched 2026-05-17: 654 stars, pushed 2026-05-12, GPL-3.0 | Android video editing baseline, Media3/Compose-relevant comparison target | Closest direct open-source Android comparator; use for feature coverage and mobile interaction expectations. |
+| LosslessCut | Desktop smart-cut editor | GitHub metadata fetched 2026-05-17: 40,487 stars, pushed 2026-05-10, GPL-2.0 | Stream-copy, fast trim, low-loss export workflows | Validates NovaCut's smart-render and mixed copy/re-encode roadmap priority. |
+| Kdenlive | Desktop NLE | GitHub metadata fetched 2026-05-17: 5,055 stars, pushed 2026-05-17, GPL-3.0 | Pro timeline, effects, proxy workflows, keyframes, titles, color, audio tools | Benchmark for timeline depth, proxy/media management, and non-destructive editing semantics. |
+| Shotcut | Desktop NLE | GitHub metadata fetched 2026-05-17: 13,963 stars, pushed 2026-05-17, GPL-3.0 | MLT-backed editing, filters, transitions, export control | Benchmark for effect/filter organization and export surface. |
+| OpenShot | Desktop NLE | GitHub metadata fetched 2026-05-17: 5,769 stars, pushed 2026-05-16 | General-purpose NLE, transitions, titles, animation | Benchmark for approachable project setup and timeline editing concepts. |
+| OpenTimelineIO | Interchange format/library | GitHub metadata fetched 2026-05-17: 1,864 stars, pushed 2026-05-01, Apache-2.0 | Timeline interchange, schema-driven project exchange | Confirms that import/export interoperability should stay a roadmap pillar. |
+| Gyroflow | Stabilization tool | GitHub metadata fetched 2026-05-17: 8,758 stars, pushed 2026-05-16, GPL-3.0 | Gyro/lens-aware stabilization | Prefer sidecar/project import and algorithm reference before attempting a full in-app gyro stack. |
+| gl-transitions | Transition registry | GitHub metadata fetched 2026-05-17: 2,085 stars, pushed 2026-05-03 | GLSL transition ecosystem | Useful template/effect-market compatibility reference. |
+
+## Commercial Products
+
+| Product | Type | Public Positioning/Feature Signal | NovaCut Lesson |
+|---|---|---|---|
+| CapCut | Mobile/desktop creator editor | Templates, AI-assisted editing, social-first workflows, asset ecosystem | Do not compete only on feature volume. Compete with local-first privacy, transparent models, and deterministic export. |
+| DaVinci Resolve | Professional editor | Professional editing, color, Fusion, Fairlight, AI-assisted workflow features in recent releases | Use as the high-end benchmark for multicam, captions, keyframes, assistant editing, and export trust. |
+| PowerDirector | Prosumer editor | AI effects, templates, mobile/prosumer export workflows | Supports prioritizing one-tap presets and polished assistant flows. |
+| LumaFusion Android | Pro mobile editor | Mobile-first pro timeline, multi-track editing, export control | Strong benchmark for mobile ergonomics and pro-feature density without desktop complexity. |
+| KineMaster | Mobile creator editor | Templates/assets, layer editing, creator workflow | Useful reference for template browsing, quick edits, and media picker affordances. |
+| VN Video Editor | Mobile creator editor | Lightweight mobile editing, creator-friendly workflows | Reference for low-friction editing and simple export flow. |
+| Adobe Premiere Rush | Retired mobile/creator editor | Official end-of-life source found | Market signal: cross-platform mobile editors are hard to sustain; NovaCut should avoid fragile cloud service dependency. |
+
+## Feature Pattern Matrix
+
+| Pattern | Commercial Presence | OSS Presence | NovaCut State | Roadmap Implication |
+|---|---|---|---|---|
+| Templates and assets | Strong in CapCut, KineMaster, PowerDirector | Emerging in OpenCut, registries like gl-transitions | Template/plugin metadata exists; marketplace/UI incomplete | Keep C.15, but tie it to local/offline and compatibility metadata first. |
+| AI auto-edit | Strong marketing trend in CapCut/DaVinci/PowerDirector | Thin in OSS mobile | Cut Assistant and transcript-driven scaffolds exist | Extend reversible proposals and review UI before freeform prompt editing. |
+| Captions/subtitles | Common across commercial apps | Common in desktop NLEs | Strong caption roadmap and shipped karaoke/locale scaffolds | Add translation/evaluation and AD audio export after model registry closure. |
+| Smart render / stream copy | Expected in LosslessCut and pro export tools | Strong in LosslessCut | Whole-timeline stream copy and planner scaffolds exist | Finish concat/per-run composer after FFmpeg decision. |
+| Diagnostic/support bundles | Common in mature media tooling | Varies | Engine exists; UI missing | R7.1 is a high-trust, low-risk priority. |
+| Interchange | Pro tools and OTIO ecosystem | Strong through OTIO/Kdenlive/Shotcut | Export exists; import harder | Preserve C.14 import as a meaningful pro differentiator. |
+| Stabilization | Common commercial feature | Gyroflow is strong OSS reference | Optical/gyro roadmap exists | Start with Gyroflow sidecar import before custom gyro implementation. |
+| Local privacy | Weak in cloud-first commercial tools | Strong in OSS but uneven UX | Core NovaCut philosophy | Make this visible in model-management and diagnostics UX. |
+
+## Positioning Conclusion
+
+NovaCut should avoid becoming a generic clone of commercial creator editors. The winning path is a narrower but deeper Android-native editor:
+
+1. Offline-first professional editing.
+2. Transparent local model delivery.
+3. Reversible assistant suggestions.
+4. Reliable export and diagnostics.
+5. Practical interchange with desktop/pro workflows.
+6. User-visible privacy guarantees.
+
+This maps directly to the Round 7 Now priorities: diagnostic export UI, model checksum closure, dependency stabilization, FFmpeg 16 KB/license decision, and Media3 Lottie parity testing.

.ai/research/2026-05-17/DATASET_MODEL_INTEGRATION_REVIEW.md

@@ -0,0 +1,93 @@
+# Dataset, Model, and Integration Review - 2026-05-17
+
+## Scope
+
+NovaCut has a meaningful data/model/integration surface: ASR, captions, translation, denoise, matting, segmentation, frame interpolation, upscaling, templates, stock assets, export/interchange, and optional cloud providers. This file focuses on what should be activated, deferred, measured, or governed.
+
+## Current Local Model Governance
+
+Evidence:
+
+- `docs/models.md` is the canonical model registry.
+- `ModelDownloadManager` exists in source.
+- The roadmap repeatedly ties large models to explicit downloads, Play Asset Delivery, and F-Droid-compatible variants.
+
+Strengths:
+
+- The project already treats models as governed artifacts, not hidden assets.
+- Licensing, source URL, delivery channel, and privacy posture are explicit concerns.
+- The local-first stance is coherent with the product positioning.
+
+Current gap:
+
+- Some model rows still contain unresolved checksum/source/activation details. No new model activation should proceed until checksums and delivery policy are closed.
+
+## Candidate Model Families
+
+| Area | Candidates | Local State | Recommendation |
+|---|---|---|---|
+| ASR / transcription | Moonshine, Whisper, Sherpa-ONNX, whisper.cpp references | ASR/caption roadmap and model policy exist | Add WER/speed/device-tier evaluation before expanding model options. |
+| Caption translation | MADLAD-400, Bergamot-style mobile translation | Caption translation data model exists | Treat as P1 after model registry closure and phrase-level evaluation fixtures. |
+| Noise reduction | DeepFilterNet 3, AndroidDeepFilterNet reference | Noise reduction engine/stub path exists | Good P1 activation once native/model integrity is verified. |
+| Stem separation | Demucs | Roadmap recognizes high complexity | Defer until audio pre/post pipeline and package size budget are clear. |
+| Matting | Robust Video Matting | Model roadmap exists | Defer until PAD and evaluation harness. |
+| Segmentation | SAM 2.1 Hiera Tiny, MediaPipe Image Segmenter | Tap-to-segment roadmap exists | Keep SAM 2.1 as watch/activation target; use MediaPipe where practical for mobile baseline. |
+| Frame interpolation | RIFE / NCNN / Vulkan | Roadmap item exists | Later. Native/Vulkan/device-tier risk is high. |
+| Upscaling | Real-ESRGAN | Roadmap item exists | Later. Large model and performance risk. |
+| Text/prompt assistant | Gemini Nano / ML Kit GenAI Prompt API | Watch item | Keep optional and local-first; no core dependency on device availability. |
+
+## Integration Candidates
+
+| Integration | Value | Risks | Recommendation |
+|---|---|---|---|
+| Play Asset Delivery | Keeps base app size manageable for large models | Play-only; F-Droid divergence | Necessary before more large model bundles. |
+| F-Droid model flavor | Preserves open/offline distribution posture | Requires alternative download/build policy | Keep explicit in `docs/models.md`. |
+| Stock asset providers | Expands media picker | Network, provider licensing, privacy | Defer until provider consent and license display are designed. |
+| Cloud AI providers | Enables heavy generative video/lip-sync | Media upload privacy, provider cost, account state | Explicit opt-in only; never default. |
+| OpenTimelineIO/FCPXML | Pro interoperability | Parser fidelity and project mapping complexity | Good P1/P2 after export pipeline stabilizes. |
+| Gyroflow sidecar/project import | Strong stabilization capability without full reimplementation | File format and sensor sync complexity | Prefer import/reference path before custom gyro engine. |
+| FFmpeg 16 KB package | Unlocks export features | License, native compliance, fork trust | P0/P1 decision gate. |
+
+## Evaluation Harness Opportunities
+
+| Area | Metric/Fixture | Why |
+|---|---|---|
+| ASR | WER/CER on short clips, noisy clips, multilingual clips | Prevents anecdotal model choice. |
+| Caption timing | Word-boundary alignment error, subtitle overlap checks | Protects karaoke/translation/edit UX. |
+| Denoise | PESQ/STOI or proxy metrics plus listening fixtures | DeepFilterNet activation needs objective and subjective checks. |
+| Segmentation/matting | IoU, boundary F-score, matte temporal stability | Prevents impressive still-frame demos that fail video use. |
+| Frame interpolation | Temporal artifact review, dropped-frame count, speed by device tier | RIFE-like features are device-sensitive. |
+| Upscaling | PSNR/SSIM where useful plus visual regression frames | Avoids sharpening artifacts and hallucinated detail. |
+| Export/render | Golden frame diffs, audio duration drift, container metadata | Needed for FFmpeg, Lottie, blend mode, HDR changes. |
+| Model downloads | Checksum failure, interrupted download, resume, offline state | Makes model management trustworthy. |
+
+## Packaging and Distribution Notes
+
+Model activation should follow this sequence:
+
+1. Add exact model source URL, license, size, SHA-256, and version.
+2. Decide delivery channel: bundled, Play Asset Delivery, explicit download, F-Droid excluded, or cloud-only.
+3. Add checksum validation tests.
+4. Add UI state for missing, downloading, failed, verified, and disabled models.
+5. Add at least one evaluation fixture before exposing the feature as active.
+
+Avoid:
+
+- Bundling unpinned model binaries.
+- Quiet network downloads.
+- Cloud fallbacks that upload media without explicit confirmation.
+- Treating benchmark results from flagship devices as representative of all supported Android hardware.
+
+## Recommended Next Model/Integration Work
+
+1. Close `docs/models.md` checksum/source/license gaps.
+2. Add model registry contract tests.
+3. Add model download failure/retry UI states.
+4. Choose PAD/F-Droid channel rules for each model family.
+5. Activate DeepFilterNet 3 only after the above is complete.
+6. Add ASR evaluation fixtures before expanding from existing transcription paths.
+7. Keep SAM/RIFE/Real-ESRGAN/RVM as later device-tier work.
+
+## Why This File Is Not Thin
+
+NovaCut has significant AI/ML/search/integration relevance. The project already contains model-management architecture, ML dependency choices, caption/transcription features, media import/export integrations, and planned cloud/on-device feature paths. The main need is governance, evaluation, and user-visible trust work before additional model activation.