winnow/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.3.0] - 2026-06-13

Added

• get_tracked_frigate_filenames() — new upload-tracker function that returns the set of Frigate filenames currently mapped for a person. Used internally as a reconciliation baseline when the Frigate GET endpoint is unreachable; also available to callers that need the mapped filename set without a count.
• Community scaffolding: CONTRIBUTING.md, SECURITY.md, GitHub issue templates (bug report, feature request), and pull request template.
• OCI image labels: org.opencontainers.image.* labels added to the runtime stage of the Dockerfile so image metadata is surfaced by container registries.
• Additional tracker tests: coverage added for get_tracked_frigate_filenames and for get_lowest_quality_mapped_file with the exclude parameter.

Fixed

• Quality score scale mismatch (USE_FULL_RESOLUTION=true): the time-spread quality-score fallback called assess_quality on the full-resolution download, while the embedding path always scores on preview thumbnails. Laplacian variance scales with image resolution, so the two paths produced incomparable scores for people with mixed-mode files. The fallback now caps the image at 1440 px before scoring to match the thumbnail scale.
• assess_quality failure left file permanently unreplaceable: if assess_quality raised an exception (e.g. an RGBA image with an unsupported channel count), score_map kept None and mark_uploaded(score=None) skipped writing the score. The uploaded file was then permanently invisible to get_lowest_quality_mapped_file because it had no entry in scores{}. The fallback now converts the image to RGB before scoring and stores 0.0 on any exception, so every uploaded file is eligible for future quality replacement.
• Uploads during Frigate GET outage never mapped: when GET /api/faces failed at upload start, the reconciliation guard (_snapshot is not None) correctly skipped the post-upload diff — but uploads that succeeded during the outage were never recorded in frigate_files, causing get_tracked_frigate_file_count to permanently under-report and Frigate to eventually exceed MAX_AUTO_IMAGES. The code now uses the tracker's mapped filenames as a pre-upload baseline when the live snapshot is unavailable, so reconciliation proceeds normally (the >target guard handles concurrent external uploads as before).
• Freed quality-replacement slot could be filled by a worse image: when a replacement delete succeeded but the subsequent upload failed all retries, effective_count stayed decremented and the next file in the iteration uploaded unconditionally — it could have a lower quality score than the file that was deleted. A min_quality_score_for_slot variable now records the deleted file's score on a successful delete; any candidate that doesn't beat that floor is skipped until the slot is filled by a qualifying image or the run ends.
• CI multi-arch and lockfile bot: hardened the build workflow — lockfile-update bot no longer races against Docker publish on the same push event; CPU image now builds for both linux/amd64 and linux/arm64; paths-ignore prevents documentation-only pushes from triggering image builds.
• README pipeline diagram updated: step 8 now explicitly documents the quality-replacement decision tree (below cap → upload, at cap + enabled → swap if better, at cap + disabled → skip).

[0.2.13] - 2026-06-13

Added

• Quality replacement: when a person is at MAX_AUTO_IMAGES, winnow now checks each new candidate against the lowest-quality image already in Frigate and swaps it in if the new image scores higher. Only images winnow uploaded (tracked in frigate_files) are ever replaced — files added manually through Frigate's UI are left untouched permanently. Enabled by default; set QUALITY_REPLACEMENT=false to revert to the previous behaviour of skipping people at cap.
• Frigate filename mapping: each successful upload now records the mapping from Frigate's assigned filename to the originating Immich asset ID and face confidence score in the tracker (frigate_files field). This is the foundation for quality replacement and future management of the Frigate training set.
• QUALITY_REPLACEMENT env var (default true): controls whether at-cap people are eligible for quality replacement. When disabled, people at MAX_AUTO_IMAGES are skipped as before.
• NOTICES file: third-party attribution for if_curator (MIT, Copyright © 2026 Sebastian) added to satisfy upstream license requirements.

[0.2.12] - 2026-06-13

Added

• ROCm (AMD GPU) support: new :rocm image tag. InsightFace runs via ROCmExecutionProvider; SigLIP runs via PyTorch ROCm 6.3 (ROCm builds expose torch.cuda.is_available() == True, so the existing CUDA path is reused automatically). Requires /dev/kfd and /dev/dri device passthrough plus video and render group membership — see compose.yml for the snippet.
• Intel GPU support: new :intel image tag. InsightFace runs via OpenVINOExecutionProvider from onnxruntime-openvino. By default OpenVINO targets CPU (no device passthrough needed); set OPENVINO_DEVICE=GPU to target Intel Arc discrete or integrated graphics. Intel's GPU compute runtime (Level Zero + OpenCL ICD) is installed automatically from Intel's official graphics repo in the image — no manual package installation required. SigLIP uses CPU inference for now (Intel Extension for PyTorch has no Python 3.13 wheels yet; the torch.xpu path is wired and will activate automatically when they ship).
• OPENVINO_DEVICE env var: controls the OpenVINO execution provider device for the :intel variant. CPU (default) requires no device passthrough. GPU targets Intel Arc discrete and integrated graphics via Level Zero.
• AMD and Intel device passthrough snippets in compose.yml: documented as commented-out alternatives to the NVIDIA deploy: block.
• :rocm and :intel CI jobs: docker-publish.yml now builds and pushes :rocm / :dev-rocm and :intel / :dev-intel alongside :latest and :cpu. release.yml builds all four variants on tag push.

Fixed

• Interactive custom-count prompt firing for all choices: in the no-embedding fallback path of the strategy selector, IntPrompt.ask was inside a dict literal and evaluated eagerly — users selecting Standard (30) or Broad (100) were still prompted to enter a custom image count. Each choice is now handled in a dedicated branch.

[0.2.11] - 2026-06-12

Fixed

• GPU broken on x86_64 Linux: insightface 1.0.1 (pulled in by the 0.2.10 lock update) added a hard dependency on the CPU onnxruntime package. Combined with an incorrect override-dependencies entry introduced in 0.2.10, both onnxruntime (CPU) and onnxruntime-gpu were being installed into the same venv. The CPU package landed last and overwrote the GPU one, causing CUDAExecutionProvider to disappear from the provider list even when a GPU was present. Fixed by declaring onnxruntime and onnxruntime-gpu as conflicting packages in uv's resolver, ensuring only the correct one is installed per platform.
• OOM crash on large person libraries (CPU mode): All candidate thumbnails were downloaded into a single in-memory dict before any processing began. At ~5 MB per decoded preview image, a person with 472 candidates would accumulate ~2.4 GB of thumbnail data alone, exhausting a 4 GB container memory limit. Thumbnails are now downloaded and processed in batches of 32, with each image released immediately after embedding. Peak in-flight thumbnail memory is now bounded to ~256 MB regardless of candidate pool size. GPU users also benefit from lower host RAM pressure and faster time-to-first-result on large libraries.

[0.2.10] - 2026-06-12

Added

• VERBOSE env var: set VERBOSE=true to enable DEBUG-level console output. The log file always captures DEBUG; this flag controls what appears on the terminal. Useful when diagnosing issues without a full shell into the container.
• :cpu Docker image tag: a separate CPU-only image (ghcr.io/sudolulo/winnow:cpu) is now built and pushed alongside :latest. Uses onnxruntime instead of onnxruntime-gpu; ~2 GB smaller. Suitable for systems without an NVIDIA GPU.
• Empty CRON_SCHEDULE keeps container alive: setting CRON_SCHEDULE= (empty string) starts the container without running immediately and without exiting — useful for docker exec ad-hoc runs on a long-lived container. Previously, an empty value was treated the same as unset (run once, then exit).

Changed

• TTY auto-detection replaces AUTO_MODE: winnow now detects whether a TTY is attached (sys.stdin.isatty()) and switches between interactive and auto mode automatically. AUTO_MODE=true becomes an explicit override for forcing auto mode in a terminal session. No config change needed for normal Docker deployments.
• Logging levels audited: internal algorithmic detail (clustering steps, asset fetch progress, per-page pagination) demoted from INFO to DEBUG. INFO now reflects meaningful pipeline milestones only (model ready, selection complete, quality filtered). Reduces noise in production logs without losing information.
• Model load logging improved: SigLIP and InsightFace loading now reports cache hit/miss, download size estimate, device used, and load time.

Fixed

• GPU OOM crash loop (production): CUDAExecutionProvider was silently absent even with a GPU attached, causing InsightFace to run on CPU and exhaust RAM processing large person libraries. Root cause: CUDA/cuDNN libraries in nvidia pip packages were invisible to onnxruntime. Fixed by running ldconfig over all nvidia-*/lib/ directories in the venv at image build time.
• ldconfig path now Python-version-agnostic: the find command used to register nvidia pip libraries hardcoded python3.13; replaced with python3.* glob so the path survives a Python upgrade without silently producing an empty ldconfig config.
• PYTHONPATH=/app added to Dockerfile: the entry point script sets sys.path[0] to the script directory, not /app. Since uv sync runs before COPY winnow/, the wheel has only dist-info in site-packages. PYTHONPATH=/app makes the winnow package importable without reverting to python -m.
• CPU fallback retrying broken GPU provider: InsightFace CPU fallback omitted providers=["CPUExecutionProvider"], causing onnxruntime to retry CUDAExecutionProvider on every inference call. Now explicitly sets the CPU provider and suppresses C-extension noise via fd-level redirect.
• _suppress_output stderr loss on fd exhaustion: if the first os.dup2 in the finally block raised OSError, the second call was skipped, permanently redirecting stderr to /dev/null for the process lifetime. Wrapped in nested try/finally so both restores are always attempted.
• Frigate /api/faces response parsing: the response is {person_name: [files], "train": [...]} — "train" is a flat pending list, not a person. Previous code called .items() on the "train" value (a list), crashing with AttributeError. Now skips the "train" key explicitly.
• Immich 401 detection: a stale or invalid API key now logs a clear error message (Immich API key is invalid or expired (401 Unauthorized)) instead of raising an unhandled exception.
• Falsy-zero detection confidence: face.get("score") or face.get("confidence") treated a valid score=0.0 as falsy, falling through to the confidence field (often None). Replaced with an explicit None check. Affected both quality filtering and hard-example weighting in diversity selection.
• Face crop using wrong person's image dimensions: in multi-person assets, _crop_face_from_thumbnail's scale-factor loop matched the first person with any face regardless of person_id, producing incorrectly scaled bounding box coordinates for the target person. Loop now applies the same person_id filter as _get_face_bbox.
• Adaptive stopping bypassed for partially-trained people: in auto mode with already_uploaded > 0, limit was converted from "auto" to an integer, disabling the FPS adaptive threshold and early-stop check. Now keeps limit="auto" through selection and trims the result to the remaining capacity afterward.
• Embedding cache key mismatch: HuggingFace cache path check hardcoded the model slug string; replaced with a derivation from model_name using "models--" + model_name.replace("/", "--") so the check stays correct if the model name changes.
• Scheduler: sleep until next run: the loop slept a fixed 60 seconds regardless of schedule interval, causing runs to fire up to 59 seconds late and waking the process unnecessarily on long schedules (e.g. weekly). Now sleeps exactly until next_run.
• Scheduler swallowing SystemExit: except BaseException in the run wrapper was replaced with except Exception (with KeyboardInterrupt re-raised above), so sys.exit() calls propagate correctly.
• Log handler leak: setup_logging now closes and removes existing handlers before adding new ones, preventing file handle accumulation across repeated calls.
• RETRY_REJECTED silently applied in interactive mode: the env var was applied unconditionally even in interactive sessions. Now used only as the default for the interactive prompt so users can override it per-run.
• compose.yml comment inverted: a comment stated -it forces non-interactive mode; corrected to reflect that -it allocates a TTY (interactive mode).

Security

• API key is no longer stored in any config file. All authentication uses environment variables or .env only.

[0.2.9] - 2026-06-12

Fixed

• arm64: onnxruntime-gpu has no arm64 wheels: onnxruntime-gpu only publishes manylinux_2_27_x86_64 and manylinux_2_28_x86_64 wheels — uv sync on arm64 failed with exit code 2. Gated onnxruntime-gpu behind sys_platform == 'linux' and platform_machine == 'x86_64'; arm64 and non-Linux installs now get the CPU onnxruntime package instead.
• uv lockfile now covers arm64: Added required-environments to [tool.uv] so the lockfile is solved for both linux/x86_64 and linux/aarch64, preventing silent resolution gaps for the non-build platform.

[0.2.8] - 2026-06-12

Fixed

• Dockerfile: use add-apt-repository with isolated GNUPGHOME: the curl | gpg --dearmor approach was silently failing — gpg exits 0 on bad/empty input, leaving an invalid keyring and causing apt to skip the deadsnakes PPA entirely. Reverted to add-apt-repository ppa:deadsnakes/ppa with GNUPGHOME=$(mktemp -d) to prevent gpg from touching any pre-existing agent socket (the original QEMU crash cause).
• arm64 base: removed --platform=$BUILDPLATFORM: the Ubuntu 24.04 base for arm64 is now pulled for the target platform, so the arm64 image contains real arm64 binaries rather than amd64 binaries in an arm64 manifest.

Changed

• License changed from MIT to AGPLv3+: source and network use of winnow now require derivative works to be open-sourced under the same terms.

[0.2.7] - 2026-06-12

Fixed

• arm64 base reverted to Ubuntu 24.04: Ubuntu 26.04 ships Python 3.14, not 3.13 — python3.13 was not locatable in its repos. Reverted arm64 base to ubuntu:24.04.
• Deadsnakes PPA now uses curl+gpg instead of add-apt-repository: add-apt-repository spawns a gpg-agent which crashes under QEMU (arm64 CI). The PPA is now added by fetching the key via curl and piping through gpg --dearmor — no agent, works on both architectures.
• PPA conditional removed: both amd64 (Ubuntu 22.04 CUDA) and arm64 (Ubuntu 24.04) now go through the same deadsnakes install path, eliminating the per-arch branching and the ARG TARGETARCH dependency in RUN commands.

[0.2.6] - 2026-06-12

Fixed

• Dockerfile: ARG TARGETARCH re-declared in each stage: Docker's automatic platform ARGs are only in scope for FROM instructions, not RUN commands. The $TARGETARCH variable in the deadsnakes PPA conditional was silently empty, so the PPA was never added and python3.13 could not be located on the Ubuntu 22.04 CUDA base. Adding ARG TARGETARCH at the top of both the build and runtime stage bodies fixes the amd64 build.

[0.2.5] - 2026-06-12

Changed

• CUDA base upgraded to 13.3.0: amd64 base image bumped from nvidia/cuda:12.9.2-cudnn-runtime-ubuntu22.04 to nvidia/cuda:13.3.0-cudnn-runtime-ubuntu22.04. Python packages (torch, onnxruntime-gpu) still target CUDA 12.6 via pip-installed libraries; the base image upgrade requires a host GPU driver that supports CUDA 13+.
• torchvision>=0.27.0 added as explicit dependency: routed through the pytorch-cu126 index for linux/x86_64, ensuring it resolves to 0.27.0+cu126 (paired with torch 2.12.0+cu126) rather than being pulled from PyPI where the older 0.21.0 wheel would downgrade torch to 2.6.0.
• torch>=2.12.0 floor raised from 2.6.0 to prevent silent downgrade.
• nvidia-cudnn-cu12>=9.0.0 floor kept intentionally loose — torch pins an exact cuDNN ABI version (9.10.2.21) and must own that constraint.

[0.2.4] - 2026-06-12

Changed

• Python 3.13 across all platforms: bumped from 3.12 to 3.13. amd64 installs Python 3.13 from the deadsnakes PPA on the Ubuntu 22.04 CUDA base; arm64 uses the Python 3.13 package available natively in Ubuntu 26.04. All confirmed dependencies (insightface 1.0.1, onnxruntime-gpu 1.26.0, torch 2.12.0+cu126) have Python 3.13 wheels.
• arm64 base: Ubuntu 26.04: Python 3.12 was removed from Ubuntu 26.04's default repos; upgrading the base pulls Python 3.13 without a PPA.
• requires-python = ">=3.13" and target-version = "py313" updated in pyproject.toml.
• CI updated to Python 3.13: test.yml and update-lockfile.yml now install Python 3.13.
• uv.lock regenerated for Python 3.13.5.

[0.2.3] - 2026-06-12

Fixed

• Clear-text API key storage: config.py no longer writes API_KEY to .immich_config.json. The key must come from an environment variable or .env file. Interactive mode now prints a tip directing users to .env. Resolves CodeQL py/clear-text-storage-sensitive-data.

Changed

• CI workflow permissions: test.yml and lint.yml now declare permissions: contents: read, following least-privilege principle and resolving actions/missing-workflow-permissions scanner alerts.
• CI lockfile race condition: removed the verify-lockfile pre-job from docker-publish.yml and release.yml. The update-lockfile.yml bot maintains the lockfile; the verify step raced against it on the same push event and caused false failures. release.yml now runs uv lock inline so tag-triggered builds are always self-consistent.
• Docs moved to wiki: docs/ folder removed from the repository. Setup, Troubleshooting, and FAQ pages are now at the GitHub wiki.

[0.2.2] - 2026-06-12

Added

• Confidence scores in upload tracker: Immich face confidence scores are now stored per asset in frigate_uploaded_ids.json under by_person[name].scores. Lays the groundwork for future replacement logic (remove low-confidence uploads when better images are found).
• Frigate-authoritative capacity tracking: At startup, GET /api/faces is queried on the Frigate host to retrieve the actual number of trained images per person from the train directory (pending/unclassified queue is excluded). This count is stored as frigate_count in the tracker JSON so it survives Frigate downtime.
• Lifetime cap uses Frigate count: MAX_AUTO_IMAGES is now enforced against Frigate's live training image count rather than the local uploaded-asset tally. Fallback priority: live Frigate API → last cached frigate_count in JSON → local uploaded count.
• Startup summary shows Frigate count: Tracker summary at startup now includes the last known Frigate training count per person (e.g. 78 uploaded, 2 rejected, 42 in Frigate).
• winnow/frigate_api.py: new module encapsulating Frigate API helpers; currently exposes get_frigate_face_counts().

Changed

• upload_tracker.py: by_person entries migrated from flat list to {asset_ids, scores, frigate_count} dict. Old list format is read and migrated transparently on first write.
• mark_uploaded() now accepts an optional score keyword argument.
• get_person_summary() now returns frigate_count and scores fields alongside uploaded and rejected.

[0.2.1] - 2026-06-12

Fixed

• Container startup reinstalling packages: entrypoint.sh used uv run, which performs a sync check on every startup and re-downloaded ruff and rebuilt the package each time. Replaced with direct .venv/bin/python calls to skip the sync entirely.
• InsightFace double models/ path: INSIGHTFACE_HOME=/models caused InsightFace to download Buffalo_L to /models/models/buffalo_l (InsightFace always appends models/ to the root). Updated default in compose.yml and .env.example to /models/.insightface.
• Lint errors in CI: unused imports in tests/test_config.py and tests/test_upload_tracker.py, unsorted imports in scheduler.py — all would have failed the ruff CI check.

[0.2.0] - 2026-06-12

First release of winnow. Forked from if-curator by Sebastian and rewritten for headless Docker deployment.

Added

Headless operation

• AUTO_MODE env var — runs without any interactive prompts; required for Docker/cron use
• DRY_RUN env var — previews selection without downloading, cropping, or uploading anything
• RETRY_REJECTED env var — re-attempts assets previously rejected by Frigate's face API
• RESET_PERSON env var — clears upload and rejection history for one named person

Docker and scheduling

• Dockerfile — multi-stage build (CUDA 12.9 on amd64, plain Ubuntu on arm64); runtime stage excludes build tools (g++, python3.12-dev, curl, gnupg)
• compose.yml — fully annotated with inline comments grouped by concern
• entrypoint.sh — runs the tool once on startup, then hands off to the scheduler if CRON_SCHEDULE is set
• scheduler.py — in-process cron scheduler that keeps the container (and loaded models) alive between runs
• CRON_SCHEDULE env var — standard cron expression for recurring runs; unset exits after first run
• .dockerignore — keeps .venv, __pycache__, test files, and logs out of the image context
• Multi-arch image: linux/amd64 and linux/arm64 built and merged into a single manifest on GHCR
• tini as PID 1 init process for correct signal handling
• Non-root container user (appuser, uid 568)
• HEALTHCHECK in Dockerfile

Object mode

• TRAINING_MODE=object — runs YOLOv9c detection on full images, crops each detected instance of a target class, and saves crops to the output volume (Frigate has no training API for objects — crops are placed manually)
• OBJECT_CLASS env var — target YOLO class label (e.g. dog, cat, car); defaults to dog

People filtering

• ONLY_PEOPLE env var — comma-separated whitelist; only these people are processed
• SKIP_PEOPLE env var — comma-separated list; these people are skipped
• MIN_FACE_COUNT env var — skip people with fewer than N tagged assets in Immich
• YEARS_FILTER env var — ignore assets older than N years (default: 10)

Image quality controls (previously hardcoded)

• BLUR_THRESHOLD env var — Laplacian variance threshold for blur rejection
• MIN_CONFIDENCE env var — minimum Immich face detection confidence
• MAX_AUTO_IMAGES env var — hard cap on auto-diversity selection
• FACE_MARGIN env var — padding around bounding box crops as a fraction of face size
• USE_FULL_RESOLUTION env var — download full-res originals vs preview thumbnails
• ENABLE_FACE_ALIGNMENT env var — align to ArcFace 112×112 format via InsightFace landmarks

GPU and model configuration

• FORCE_CPU env var — disable GPU; fall back to CPU for embedding computation
• INSIGHTFACE_HOME env var — controls model persistence for Buffalo_L
• HF_HOME env var — HuggingFace model cache path for SigLIP
• LD_LIBRARY_PATH set in the image to expose CUDA and cuDNN pip libraries so onnxruntime-gpu can find them at runtime

Caching and upload tracking

• ENABLE_CACHE / CACHE_DIR env vars — opt-in embedding cache to skip recomputation on reruns
• Per-person upload tracker persisted as JSON; prevents the same asset from being uploaded twice across runs weeks apart, even if the container is recreated
• Startup summary showing uploaded and rejected counts per person
• LIMIT env var — exact image count overriding STRATEGY preset

CI/CD

• docker-publish.yml — builds multi-arch image and pushes to GHCR on push to main (:latest) or dev (:dev)
• release.yml — triggered by v* tags or workflow_dispatch; creates a GitHub Release, extracts changelog notes, builds and pushes versioned image to GHCR
• lint.yml — runs Ruff on push/PR to main and dev
• test.yml — runs pytest on push/PR to main and dev
• update-lockfile.yml — regenerates uv.lock and commits it when pyproject.toml changes
• Dependabot: weekly grouped PRs for Python dependencies (uv ecosystem) and GitHub Actions versions

Testing

• 24 unit tests across four modules: test_config, test_immich_api, test_jobs, test_upload_tracker

Documentation

• docs/setup.md — step-by-step install and GPU passthrough guide
• docs/troubleshooting.md — common failure modes with fixes
• docs/faq.md — answers to questions new users will ask
• .env.example — copy-paste starting point with every env var and inline comments
• README rewritten: pipeline diagram, env var reference tables, scheduling behaviour, requirements

Changed

• cli.py split into three focused modules — cli.py (entry point), jobs.py (configuration and strategy resolution), executor.py (download, crop, upload)
• Dependency management replaced with uv; uv.lock pins the full transitive graph for reproducible builds
• compose.yml fully annotated; all env vars documented with inline comments
• LD_LIBRARY_PATH extended to include both cuDNN and CUDA runtime libraries

Fixed

• EXIF orientation: PIL opens JPEGs without applying rotation metadata; Immich computes face bounding boxes on orientation-corrected images, so portrait photos produced misaligned crops. ImageOps.exif_transpose() now normalizes orientation before any coordinate math.
• Model persistence: FaceAnalysis was initialized with root="~/.insightface" (hardcoded), ignoring INSIGHTFACE_HOME. Buffalo_L was re-downloaded into the container on every run instead of persisting to the mounted volume.
• Upload deduplication: upload_to_frigate() scanned the output directory with os.listdir(), picking up leftover files from previous runs and re-uploading them. Now only files created in the current run are uploaded.
• RGBA images: Images in RGBA mode raised an error when encoding to JPEG. All images are now converted to RGB before saving.
• Object mode uploads: Object mode incorrectly called the Frigate face registration API. Frigate has no API for object training data — object mode now only saves crops to disk.
• Stale output files: The output directory was not cleaned between runs, causing crops to accumulate. Now wiped at the start of each face-mode run.
• JSON decode errors: get_people() and fetch_all_assets() only caught RequestException, leaving JSONDecodeError unhandled on non-JSON 200 responses from Immich.
• Spaces in names: People names with spaces caused downstream errors.
• Inconsistent headers: Some API calls used a raw header dict instead of get_headers().
• Docker layer caching: uv sync was placed after COPY if_curator/, so any source change invalidated the 800 MB dependency cache. Dependencies are now installed before source is copied.

Security

• CUDA base image bumped from nvidia/cuda:12.6.3 to nvidia/cuda:12.9.2-cudnn-runtime-ubuntu22.04, picking up Ubuntu security patches flagged by Dependabot.