YA-WAMF 2.15.0

YA-WAMF 2.15.0

YA-WAMF 2.15.0 makes first-time setup and background work easier to understand, while strengthening
historical imports, distant-bird evidence, and hardware-aware model operation.

What's new

• See what background work is actually doing. The Jobs workspace and global progress indicator
  now use one server-owned view of queued and running video analysis, best-quality snapshots,
  full-visit clips, and backfill work. Queued work is no longer presented as already running.
• Run setup safely whenever you need it. The re-runnable Setup wizard now lives in Settings,
  preserves saved secrets during connection tests, reports section readiness honestly, and can
  optionally start a visible import of retained Frigate history.
• Validate the hardware your image can really use. Model and crop-detector checks share the
  packaged-runtime, detected-hardware, and model-support contract. The active model shows its
  actual provider and fallback order, including validated Intel OpenVINO CPU, GPU, and NPU paths
  where the selected image, host, and model support them.
• Raspberry Pi images start with a working fallback model. ARM64 builds use the smaller LiteRT
  runtime, include an offline classifier, and must complete real inference under QEMU before their
  mutable tags are published.

Smoother everyday use

• Backfill now follows the same acceptance rules as live detections. Completed Frigate
  snapshots are fetched deterministically and reconstruct their aligned tracked-object crop before
  classification. Confidence, blocked-species, abstention, and trusted-sublabel gates are shared;
  weak historical results are skipped rather than admitted through a separate rule.
• High-quality snapshots protect Frigate's strongest evidence. Model-generated crops must agree
  with the known species and improve classifier evidence before replacing a valid Frigate crop or
  full frame. Small distant birds can use a more permissive localization search only inside the
  multi-image evidence path, where full-frame and Frigate alternatives remain available.
• Bursts and restarts lose less work. Detection intake is idempotent, automatic video jobs are
  durable before entering memory, bounded queues expose their pressure, and shutdown drains MQTT
  before downstream workers stop.
• An inconclusive video no longer disables later analysis. A healthy temporal run with weak or
  conflicting frame evidence remains visible as no confident result, but cannot open the video
  circuit breaker. Genuine worker failures, exceptions, and timeouts still retain the cooldown.

Before you update

• Back up /config and /data, then use a normal image update. Two automatic database
  migrations add saved crop-strategy provenance and stable BirdNET source-event identity. No
  manual database command or configuration rewrite is required.
• Keep the unsuffixed tag unless you intentionally want a smaller runtime image. latest and
  v2.15.0 remain the full compatibility build. Read the
  hardware-acceleration guide
  before switching to v2.15.0-cpu, v2.15.0-intel, or v2.15.0-cuda; change only the image tag
  so the existing /config, /data, models, and provider preference are preserved.

Good to know

• Full hardware validation remains an explicit action. Diagnostics may download missing model
  assets and test the providers available to the running image. Container startup performs only
  the bounded checks needed to load the selected model; the longer accelerated-versus-CPU
  benchmark remains opt-in.
• Raspberry Pi support remains best-effort pending physical-device evidence. CI proves the
  ARM64 image starts and classifies under QEMU, but sustained thermal, storage, and accelerator
  testing on representative Pi hardware is still required.

See the full changelog
for the complete technical record.