Skip to content

progress.md

Latest commit

 

History

History
174 lines (171 loc) · 27.6 KB

progress.md

File metadata and controls

174 lines (171 loc) · 27.6 KB

Progress: Desktop CPU Batch OCR Evaluation

2026-05-21

  • Started batch desktop CPU throughput/accuracy evaluation work.
  • Added tests for PaddleOCR CTC decoding, throughput summary, model output layout selection, and PaddleOCR metadata parsing.
  • Implemented PaddleOCR ONNX evaluation, batch throughput measurement, and candidate config metadata paths.
  • Fixed a LightSVTR regression caused by treating [time,batch,class] output as [batch,time,class].
  • Ran all Python tests: pixi run python -m unittest discover -s tests -v passed 23 tests.
  • Ran full CPU batch report with TrOCR imported predictions: /tmp/medlog_bare_benchmark/candidate_results_cpu_batch_with_trocr.json and /tmp/medlog_bare_benchmark/candidate_results_cpu_batch_with_trocr.txt.
  • Started PARSeq integration work. Read planning-with-files and TDD guidance, then began verifying the real checkpoint/API before coding.
  • Added failing PARSeq integration tests first, then implemented export_parseq_onnx.py and evaluate_parseq_onnx.py.
  • Verified PARSeq unit tests: pixi run python -m unittest tests.test_parseq_integration -v passed 4 tests.
  • Exported PARSeq to exported_candidates/parseq.onnx plus external data file exported_candidates/parseq.onnx.data.
  • Added direct PARSeq support to run_candidate_evaluation.py, then verified integrated report: /tmp/medlog_bare_benchmark/candidate_results_cpu_batch_parseq_direct.json and /tmp/medlog_bare_benchmark/candidate_results_cpu_batch_parseq_direct.txt.
  • Implemented FastViT-T8 CTC candidate with multi-scale feature fusion, two-stage training, ONNX export metadata, and torch_ctc_onnx runner support.
  • Ran FastViT no-pretrained CPU smoke train/export: /tmp/fastvit_ctc_smoke/fastvit_t8_ctc.onnx.
  • Ran FastViT smoke candidate evaluation: /tmp/fastvit_ctc_smoke/results.json and /tmp/fastvit_ctc_smoke/results.txt.
  • Added kaggle_candidate_finetune_kernel for Kaggle-side all-candidate reporting and fine-tuning of trainable architectures.
  • Verified plan-only run: pixi run python kaggle_candidate_finetune_kernel/kaggle_candidate_finetune.py --output-dir /tmp/kaggle_candidate_plan --plan-only --candidates all.
  • Pushed Kaggle candidate fine-tune kernel version 2; remote status reported KernelWorkerStatus.RUNNING.

2026-05-22

  • Pulled Kaggle candidate fine-tune results from /tmp/medlog_kaggle_candidate_results/candidate_finetune and analyzed FastViT/LightSVTR accuracy, capacity, and local CPU latency.
  • Added failing tests first for PaddleOCR Kaggle trainable candidates and PaddleOCR config materialization.
  • Implemented PaddleOCR Kaggle dispatch support for ppocrv5_mobile_rec, ppocrv5_server_rec, repsvtr, and svtrv2_server.
  • Verified PaddleOCR candidate plan: pixi run python kaggle_candidate_finetune_kernel/kaggle_candidate_finetune.py --output-dir /tmp/kaggle_candidate_plan_paddle --plan-only --candidates all.
  • Verified PaddleOCR config materialization smoke under /tmp/kaggle_paddle_config_smoke.
  • Added Android instrumentation benchmark for FastViT ONNX Runtime CPU/NNAPI and helper script scripts/run_fastvit_nnapi_benchmark.sh.
  • Compiled Android benchmark: ./gradlew :app:compileDebugAndroidTestKotlin.
  • Ran FastViT reparameterized ONNX on connected M2012K11C: CPU mean_ms=49.70 p50_ms=50 p95_ms=51 throughput_sps=20.12; NNAPI mean_ms=49.20 p50_ms=49 p95_ms=50 throughput_sps=20.33; nnapi_cpu_disabled one-run smoke mean_ms=52.00.
  • Pushed Kaggle candidate fine-tune kernel version 3; remote status reported KernelWorkerStatus.RUNNING.
  • Probed Kaggle v3 while running: output files still reflected the previous run and logs were empty.
  • Added a failing test and fix for PaddlePaddle GPU installation, defaulting Kaggle to paddlepaddle-gpu==3.3.0 via the official CUDA 12.6 wheel index.
  • Pushed Kaggle candidate fine-tune kernel version 4 with the official CUDA 12.6 Paddle wheel index; remote status reported KernelWorkerStatus.RUNNING.
  • Added failing tests and fixes for the v4 CUDA dependency conflict: FastViT now runs before PaddleOCR, and Paddle installs with --no-deps unless --paddle-install-deps is explicitly set.
  • Pushed Kaggle candidate fine-tune kernel version 5 with FastViT-before-Paddle ordering and Paddle --no-deps; remote status reported KernelWorkerStatus.RUNNING.
  • Started formal Android package rename and debug-device install task.
  • Confirmed connected adb device: 6b9f2b84 (M2012K11C).
  • Renamed Android app/source package to com.driezy.medlog, including Gradle namespace/applicationId, Kotlin packages/imports, manifest actions, shortcuts, widget config class names, benchmark script package names, and Room schema export directory.
  • Verified ./gradlew :app:assembleDebug, ./gradlew :app:installDebug, and ./gradlew :app:compileDebugUnitTestKotlin :app:compileDebugAndroidTestKotlin.
  • Launched com.driezy.medlog/.ui.MainActivity on M2012K11C; process 21676 ran in foreground with no recent fatal logcat entries for the new package.
  • Optimized OCR settings UI before commit by extracting the repeated model option card, moving visible badge/spec labels into string resources, and cancelling SevenSegmentRecognizer's settings collection scope on close.
  • Verified optimization with ./gradlew :app:ktlintCheck and ./gradlew :app:compileDebugKotlin :app:compileDebugUnitTestKotlin :app:compileDebugAndroidTestKotlin.
  • Started Material Expressive home layout implementation slice focused on layout hierarchy, emphasized typography, surface elevation, and primary action placement.
  • Completed the home layout slice: Today overview now owns progress, streak, next-dose, and take-all action; empty state no longer duplicates the add action with a FAB.
  • Started HealthScreen Material Expressive slice focused on OCR hero action, health metric sections, supporting recent-record hierarchy, and emphasized metric typography.
  • Completed the HealthScreen slice: added a primary OCR scan hero with manual-entry secondary action, wrapped stats in a health metrics section, lowered recent records into supporting content, and separated emphasized metric values from smaller unit labels.
  • Confirmed OCR processing overlay and HealthScreen loading state already use Material3 Expressive LoadingIndicator.
  • Verified HealthScreen changes with ./gradlew :app:compileDebugKotlin, ./gradlew :app:ktlintCheck, ./gradlew :app:assembleDebug, ./gradlew :app:installDebug, and a no-fatal adb logcat check after launching com.driezy.medlog/.ui.MainActivity.
  • Completed the broader Material Expressive pass: Home now defaults to "Now" and "Later today" task groups with PRN separated, time-period group containers use flat surface hierarchy, and Settings is consolidated into the requested Appearance / Reminders / OCR & Health / Widgets / Data & About containers.
  • Re-verified with ./gradlew :app:compileDebugKotlin, ./gradlew :app:ktlintCheck, ./gradlew :app:assembleDebug, ./gradlew :app:installDebug, and launched com.driezy.medlog/.ui.MainActivity on M2012K11C with no recent fatal logcat entries.
  • Completed a Material Expressive micro-motion pass after checking current official Compose Material3 motion guidance: Home progress digits, medication status confirmation, OCR content swaps, Add Medication dynamic form fields, Settings expand/collapse groups, and shared animated list items now use MotionScheme spatial specs for movement/size and effects specs for alpha/color.
  • Installed the micro-motion build on M2012K11C and launched com.driezy.medlog/.ui.MainActivity; no recent fatal logcat entries were found.
  • Started current lint cleanup after :app:lintDebug report showed 2 errors and 18 warnings despite Gradle success due to lint abort being disabled.
  • First post-upgrade lint run failed because AGP 9.2.1 requires Gradle 9.4.1; updated the Gradle wrapper URL and checksum before retrying.
  • :app:lintDebug passed with "Lint found no errors or warnings"; parsed app/build/reports/lint-results-debug.xml and confirmed total=0.
  • Combined ktlint/test/assemble run failed once at :app:packageDebug; isolated :app:packageDebug --stacktrace then passed, so continued by suppressing the remaining JVM test warning via -Xshare:off.
  • Re-ran ./gradlew :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug; build passed cleanly with no visible warning output.
  • Re-ran ./gradlew :app:lintDebug after build-script changes; lint passed with no errors or warnings and XML lint_issues=0.
  • Installed debug APK on connected M2012K11C (6b9f2b84), launched com.driezy.medlog, confirmed process startup, and found no recent fatal AndroidRuntime entries for the app.
  • Pushed 48e81b9 Resolve Android lint warnings; GitHub CI passed ktlint, unit tests, and Android Lint, then failed at Build Debug APK with PackageAndroidArtifact$IncrementalSplitterRunnable.
  • Verified local ./gradlew clean :app:assembleDebug --stacktrace passes; updated CI and release workflows to opt JS actions into Node 24 and run clean APK builds with stacktraces.
  • GitHub CI for 1891a37 passed, but still emitted a Node 20 deprecation annotation because v4 JS actions were merely forced onto Node 24. Checked upstream tags and upgraded CI/release actions to current Node 24-capable major versions.
  • Started deterministic debug seed data and all-page Material Expressive audit request. Stopped the interrupted local gh run watch process; GitHub Release run for v1.15.5 was still in progress at handoff.
  • Added red SeedDataFormatTest, then implemented debug-only SeedDemoDataUseCase, deterministic SeedDemoCalendar, SeedDemoProfile, and DebugSeedReceiver.
  • Verified SeedDataFormatTest passes. Installed debug APK on 6b9f2b84, seeded via explicit adb broadcast with reset=true profile=standard, and queried copied Room database: medications=5, logs=8, health_records=6, today_taken_scheduled=3/7, today_taken_all=4/8, low_stock=2, health latest values for all six HealthType entries, orphan_logs=0.
  • Unified the first scanner/entry surface pass: camera permission empty state, OCR/health OCR/QR guidance pill, QR viewfinder overlay, OCR flash button, Home QR entry, Health OCR toolbar entry, Add Medication OCR entry, and QR dialog import/share/replace actions. Verified with ktlint, unit tests, assemble, install, adb seed, and copied Room database queries; no screenshot/visual check used.
  • Switched code-level Material audit away from long Android lint runs. Standardized top-level page primary actions to right-side ExtendedFloatingActionButton with labels on Home, Health, History, Drugs, and Diary; kept VibrantFloatingActionButton only inside the OCR camera surface as an in-context capture action. Flattened remaining low-priority ElevatedCard usages in PRN, Health BMI/chart, and Settings widget picker into surfaceContainerLow cards.
  • Applied title/FAB/I18N audit: Health and Diary now use collapsible LargeTopAppBar like other top-level destinations; Health adds a secondary SmallFloatingActionButton for manual entry while keeping scan as the labeled primary FAB. Added missing Japanese and Korean strings for Home groups, Health OCR hero/sections, OCR states, and OCR model settings; XML/i18n key parity checks, compile, ktlint, seed format test, and debug install all passed.
  • Clarified and corrected OCR scan-frame behavior: the previous camera frame was visual guidance only while OCR processed the full captured image. Added OcrRecognitionRegion so medication OCR and health OCR now crop the captured bitmap to the frame-shaped center region before ML Kit, seven-segment CRNN, and LCD display detection run; medication uses a wide text frame and health uses a device-display frame. Updated localized hints to say content should be inside the frame, added region unit tests, and verified compile, ktlint, assemble, tests, and debug install.
  • Added a pre-capture readiness panel for OCR camera flows with frame-fill, hold-steady, and glare-avoidance guidance. Added OcrFrameInputPreprocessor so the cropped frame image is normalized to a model-friendly long-edge range before OCR variants/model inference, with sampled luminance/contrast metrics logged for non-visual diagnosis. Verified XML/i18n parity, compile, ktlint, targeted OCR tests, assemble, and debug install.
  • Started OCR result grouping and home hierarchy refinement. Checked current official Material/Android guidance for M3 Expressive, emphasis, tonal elevation, FAB variants, top app bars, and Google Sans Flex / variable font direction.
  • Parallel Gradle verification triggered KSP cache/file-generation corruption in app/build/kspCaches/debug; stop running KSP-backed Gradle tasks in parallel and clear generated KSP cache before retrying serial verification.
  • Completed OCR result grouping by recognition source/model, strengthened low-stock warning surfaces, allowed important medication names to wrap instead of single-line ellipsizing, and removed negative tracking from display typography.
  • Verified serially with :app:compileDebugKotlin, targeted OcrRecognitionOutputTest, :app:ktlintCheck, :app:assembleDebug, :app:installDebug on 6b9f2b84, app launch, no recent AndroidRuntime/FATAL EXCEPTION, and full :app:testDebugUnitTest.
  • Started theme-layer Flex font integration. Current approach: use Compose ui-text-google-fonts and one shared MedLogFontFamily applied to baseline and emphasized typography roles.
  • Adjusted implementation to bundle Google Sans Flex TTF weights locally instead of using runtime downloadable fonts, then wired MedLogFontFamily into every baseline and emphasized TextStyle.
  • Verified Flex font integration with :app:compileDebugKotlin, targeted MedLogTypographyTest, :app:ktlintCheck, full :app:testDebugUnitTest, :app:assembleDebug, :app:installDebug on 6b9f2b84, app launch, and no recent AndroidRuntime/FATAL EXCEPTION.
  • Started widgets/notifications/data settings UX pass. Initial audit: widget previews are static vector placeholders with fake exact times; reminder settings lack an at-a-glance summary; backup/restore are plain list items despite restore being destructive.
  • Checked GitHub Actions Gradle caching strategy. Current workflows already used gradle/actions/setup-gradle, but split ktlint/tests/lint/package into separate Gradle invocations and still ran clean assemble*.
  • Updated CI and release workflows so GitHub Actions performs ktlint, unit tests, Android lint, and APK assembly in one cache-aware Gradle invocation with --build-cache --configuration-cache --stacktrace; removed clean builds and added debug APK artifact upload.
  • Verified workflow YAML parses locally and confirmed no clean assemble* command remains in .github/workflows.
  • Added a red MaterialSymbolsMigrationGuardTest, confirmed it failed on the old Compose Material Icons dependency/imports and missing XML icon set, then migrated icon usage to local Material Symbols XML drawables.
  • Generated 101 ic_symbol_*.xml drawables from Material Symbols Rounded paths and centralized rendering through explicit MedLogIcon(icon = ...); removed the incorrect fake Icon(imageVector: Int) compatibility layer and guarded against it in tests.
  • Removed material-icons-extended from Gradle dependencies and replaced all Icons.* imports/usages across navigation, scanning, home, settings, and the remaining modules with MedLogIcons.
  • Optimized build/CI behavior by increasing Gradle daemon heap to 4G with ParallelGC, keeping GitHub Actions on setup-gradle, disabling lint's expensive VectorPath check for official Material Symbols vectors, and keeping Gradle executions serial to avoid KSP cache contention.
  • Fixed all current lint warnings: debug seed receiver exported warning, OCR result count pluralization, unused strings, KTX bitmap pixel access, and external vector path analysis.
  • Verified :app:compileDebugKotlin :app:lintDebug, parsed lint XML with lint_issues=0, and ran :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug successfully; warm assemble completed in 8s and final combined verification in 19s.
  • Checked current official Google Fonts Material Symbols and Android Compose icon guidance. Android now recommends Google Font Icons / Material Symbols XML from the Android tab instead of the old Compose material-icons artifacts.
  • Confirmed the provided family download URL is a font/ligature asset path, not the runtime model MedLog wants for app icons. MedLog keeps icons as local VectorDrawable XML rendered through MedLogIcon and painterResource.
  • Added scripts/update_material_symbols.mjs to batch-regenerate the 101 Material Symbols Rounded VectorDrawable XML files and MedLogIcons.kt, with generated-source comments in every output file.
  • Added docs/material-symbols.md documenting the source, update workflow, and rules against runtime symbol fonts, androidx.compose.material.icons, material-icons-extended, and fake Icon(imageVector: Int) overloads.
  • Added a red-then-green reproducibility guard in MaterialSymbolsMigrationGuardTest; targeted guard test passed after adding the script, generated markers, and docs.
  • Verified the Material Symbols sourcing iteration with ./gradlew :app:compileDebugKotlin :app:lintDebug :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug; lint reported no errors or warnings.
  • Added failing guard coverage for the two Material Symbols enhancements: selected top-level navigation must use fill=1 variants, and large display symbols must use 40/48dp optical-size resources.
  • Extended scripts/update_material_symbols.mjs to generate 112 XML symbols total: 101 baseline 24dp icons, 6 selected fill=1 navigation icons, and 5 optical-size display icons.
  • Wired selected variants into bottom navigation, rail, and drawer destinations; wired optical-size variants into the drawer brand icon, OCR empty state, Health empty state, Drugs empty state, and the welcome confirmation moment.
  • Re-ran targeted MaterialSymbolsMigrationGuardTest; the new state/optical-size checks passed.
  • Started M3 Expressive motion physics pass from the May 2025 guidance: keep the product on MotionScheme.expressive(), use spatial tokens for movement/size and effects tokens for opacity/color, and avoid legacy duration/easing for custom motion hotspots.
  • Added failing MotionSystemGuardTest, then migrated navigation transitions, welcome pager/entry/pulse motion, History color transitions, and Medication Detail adherence color transitions to MaterialTheme.motionScheme spring tokens.
  • Verified motion changes with targeted MotionSystemGuardTest, then ran :app:ktlintCheck, targeted motion/icon guard tests, and :app:assembleDebug; installed and launched the debug APK on 6b9f2b84.
  • Android lint was attempted twice after the icon/motion work, but both runs stalled for more than 10 minutes in lint analysis/report generation. The verified compile/ktlint/tests/assemble/install path passed; lint needs a separate tooling-performance investigation.

2026-05-23

  • Started Material 3 Carousel adoption analysis for the Android app.
  • Read the planning-with-files skill, checked existing planning files, and appended a dedicated analysis phase without overwriting prior OCR/Material work.
  • Inspected current Compose screens, screenshots, horizontal-scroll sites, widget preview components, and local Material3 carousel classes.
  • Identified Health metrics and Settings widget previews as the strongest candidates, with Drugs featured categories as a secondary candidate and several task/text-heavy surfaces as explicit non-candidates.
  • Started implementation goal for Carousel optimization slice; using test-first guard coverage before production UI changes.
  • Added red CarouselAdoptionGuardTest; it failed as expected because Health metrics and Settings widget previews were still non-carousel layouts.
  • Replaced Health metrics' horizontal stat row with HorizontalUncontainedCarousel, keeping health filter chips unchanged.
  • Replaced the vertically stacked Settings widget picker cards with a WidgetPreviewCarousel using HorizontalCenteredHeroCarousel.
  • Verified red-to-green targeted guard test, :app:ktlintCheck, :app:assembleDebug, full :app:testDebugUnitTest, and git diff --check.
  • Audited dynamic color/color role implementation against current official Compose Material3 color scheme roles and local theme/UI usage. Result: dynamic wallpaper colors are implemented for Compose screens, but the static fallback role table and some custom/non-Compose surfaces are not fully aligned with latest role coverage.
  • Started dynamic color/color-role optimization goal and added red ColorRoleGuardTest covering current M3 role coverage, hardcoded/ad hoc color usage, notification tinting, and bare Glance widget theming.
  • Confirmed the new guard failed before implementation on missing expanded roles, calendarWarning, launcher notification tinting, and missing MedLogGlanceTheme.
  • Completed the main optimization slice: static light/dark schemes now include expanded M3 roles, custom visual state colors now use Material roles, notifications resolve theme/dynamic primary colors, and Glance widgets use project-owned ColorProviders.
  • Re-ran targeted ColorRoleGuardTest; it passed after one transient incremental ASM transform failure was cleared by a stacktrace rerun.
  • Verified the completed dynamic color slice with ./gradlew :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug and git diff --check.
  • Started Material Motion/Shape checklist pass from the provided spec. Treated MDC Views APIs as not applicable to the Compose UI, and focused implementation on Compose MotionScheme and Shapes.
  • Added red motion/shape guard coverage. The new motion guard failed on unthemed default transition specs; the shape guard failed on the custom shape scale.
  • Replaced remaining bare Compose transition specs in Welcome, Drugs, and Home task group expansion with MaterialTheme.motionScheme spatial/effects specs.
  • Aligned MedLogShapes to Material 3 role values: 4/8/12/16/28dp.
  • Verified the motion/shape pass with targeted guard tests and full ./gradlew :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug.
  • Started editorial treatment adoption from the provided guidance and screenshots. Selected Home's today-progress card as the appropriate showcase moment because it celebrates user completion/progress.
  • Added red EditorialTreatmentGuardTest, then implemented EditorialTypography theme tokens and exposed them through MaterialTheme.editorialTypography.
  • Added EditorialProgressMoment to the Home progress card with large animated progress type and localized completion text for Chinese, English, Japanese, and Korean.
  • Verified the targeted editorial guard with ./gradlew :app:testDebugUnitTest --tests com.driezy.medlog.ui.theme.EditorialTreatmentGuardTest.
  • Verified the completed editorial treatment slice with ./gradlew :app:ktlintCheck :app:testDebugUnitTest :app:assembleDebug.
  • Started external seven-segment/medical-device dataset acquisition goal.
  • Loaded planning-with-files, ran session catchup, and inspected seven_segment_ocr structure.
  • Chosen staging convention: seven_segment_ocr/external_datasets/ for third-party downloads, with raw archives/downloads separated from extracted/preprocessed copies.
  • Added seven_segment_ocr/external_datasets/ to .gitignore so downloaded third-party data remains in the workspace but is not accidentally committed.
  • Downloaded Kaggle sources via aria2 RPC after CLI downloads were slow: BP monitor, oximeter, and seven-segment-display YOLOv5.
  • Downloaded Hugging Face MiXaiLL76/7SEG_OCR by direct parquet LFS URL after snapshot_download stalled on the LFS file.
  • Extracted Kaggle archives and generated preprocessing outputs: preprocessed/kaggle_bp_monitor/device_bboxes.csv, preprocessed/kaggle_oximeter/device_bboxes.csv, preprocessed/kaggle_seven_segment_yolov5/yolo_manifest.csv, preprocessed/hf_7seg_ocr/images/, preprocessed/hf_7seg_ocr/labels.csv.
  • Wrote consolidated manifests: manifests/dataset_summary.json, manifests/medical_device_bboxes.csv, manifests/preview_contact_sheet.jpg.
  • Confirmed aria2 has no active/waiting downloads at the end of the acquisition pass.
  • Audited ground truth contents: BP/Oximeter provide only whole-device Pascal/VOC detection boxes; Kaggle seven-segment YOLO provides digit bbox detection labels; HF 7SEG provides synthetic sequence text labels.
  • Designed the unified OCR pipeline with Kaggle compatibility and PySide6 visualization. Added seven_segment_ocr/PIPELINE_DESIGN.md.
  • Implemented the unified seven-segment OCR pipeline core under seven_segment_ocr/pipeline/: schema parsing, task DAG validation, JSONL events, run artifact layout, image-text/VOC/YOLO dataset inspection, HF recognition normalization, VOC/YOLO detection preparation, subprocess adapters for training/eval, ONNX single-image inference wrapper, Android asset export copy step, and Kaggle script-kernel packaging.
  • Added seven_segment_ocr/pipeline_ui/ as a PySide6 shell that uses QProcess to call python -m pipeline.cli, tails events.jsonl, and displays dataset profiles/previews, metric curves, inference/evaluation artifacts, and Kaggle push/status/fetch output without importing PySide6 from the headless core.
  • Added seven_segment_ocr/pipeline_task.example.yaml covering the downloaded HF 7SEG OCR recognition data, Kaggle seven-segment YOLO data, and Kaggle BP VOC device-bbox data.
  • Added pipeline regression coverage in seven_segment_ocr/tests/test_pipeline_core.py; verified schema defaults/cycle rejection, event JSONL rows, dataset inspect/prepare, suspicious YOLO class 66 warning, runner artifacts/events, Kaggle package generation, example config loading, and the no-PySide6-import core boundary.
  • Ran local smoke pipeline: dataset.inspect -> dataset.prepare_recognition -> dataset.inspect YOLO -> dataset.prepare_detection -> VOC inspect; result completed=5 failed=0, with dataset profiles, preview contact sheets, normalized labels.csv, YOLO data.yaml, and sample_preview/warning events under runs/seven_segment_smoke/.
  • Ran Kaggle packaging smoke: generated runs/seven_segment_smoke/kaggle/kernel/ with kernel-metadata.json, pipeline_task.json, kaggle_pipeline_entry.py, and copied headless OCR modules; verified the entry script does not contain PySide6.
  • Switched PySide6 to pixi PyPI dependencies, ran pixi install, fixed Qt plugin discovery in pipeline_ui.app, and verified PipelineMainWindow can instantiate with QT_QPA_PLATFORM=offscreen.
  • Verified pixi run python -m unittest tests.test_pipeline_core -v passes and full Python test discovery passes 47 tests.
  • Upgraded the pipeline runner with default-on .task_cache.json, task_skipped events, per-task artifact ownership, partial graph selection (--target, --from-task, --resume), --no-cache, --force, and dependency-ready concurrent scheduling through ThreadPoolExecutor.
  • Added headless graph/cache/leaderboard helpers plus tests for hash invalidation, forced parent reruns invalidating downstream cache, resume/blocked task selection, independent branch scheduling, and CLI flag parsing.
  • Added PySide6 DAG dashboard components: DagView for node/edge rendering and event-driven node states, ConfigBuilderWidget for YAML-backed task/dataset editing and validated save-as, and LeaderboardWidget for cross-run metric comparison.
  • Integrated the UI widgets into PipelineMainWindow with DAG, Builder, and Leaderboard tabs, right-click node actions for target/from-task runs, node-specific stdout/stderr/event viewing, and artifact filtering by selected task.
  • Verified the upgraded implementation with pixi run python -m unittest discover -s tests -v: 56 tests passed.
  • Ran real pipeline smoke twice with --max-workers 4: first run completed 5 tasks, second run cache-skipped all 5 tasks with 0 blocked/failed; Kaggle package smoke still generated a script-only kernel without PySide6.