feat(trainer): phase-B body-type classifier — trainer job on the collector host + the classifier stage on the booth

apps/trainer (parking-trainer): inspect / train / evaluate / publish. Reads the wash
collector's SQLite + crops read-only off its volume; time split (validation = newest
slice); thin classes dropped; damped class weights; `features` mode (frozen ImageNet
backbone, on-disk feature cache, seconds to retrain) and `finetune` mode (light
augmentation). CPU-only torch from PyTorch's wheel index. ONNX export checked against
the torch model; NO model file below the validation floor (exit 3, report still written);
exit 2 = not enough labels. `evaluate` scores a shipped model on labels reviewed after
training + the unlabelled pile; `publish` PUTs a version folder to a Gitea generic package.
Light core deps; the `train` extra is heavy — CI syncs without it, torch tests skip.

apps/vision: BodyTypeClassifier (bodytype.onnx + sidecar = the preprocessing contract:
crop margin, input size, RGB 0-255, normalisation inside the graph) and
RefinedVehicleDetector over YOLOX — refines only `car` or a class the classifier trained
on, min-confidence, `detector_class` on the result; path set but no file = phase B off
without an error; a broken file is a health detail. models/bodytype.version (tracked,
empty) pins the published version the Dockerfile fetches at build (BuildKit secret;
a pin that cannot be fetched fails the build). Verified: a trainer model gives identical
probabilities inside the vision service; both images built and smoke-tested.

Delivery: parking-trainer image in build-images.yml, the `trainer` compose profile on the
collector stack (CPU, read-only data, TRAINER_OUT), commented TRAINER_OUT/PUBLISH_TOKEN in
the wash-collector stack, .dockerignore for both Python contexts, trainer deps synced in CI.

Wiki: bodytype-classifier-training rewritten as built (+ one fleet model not per site,
secrets/access, where the crops live), opencv-anpr-service §Phase B, vision-review-outbox,
vision-service-packaging, fleet-deployment-komodo, index, log.

Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU

.gitea/workflows/build-images.yml

@@ -1,6 +1,6 @@
 name: Build & push images
 
-# Build the SERVER (API + SPA), COLLECTOR (wash review) and VISION (ANPR) container images and push them to the
+# Build the SERVER (API + SPA), COLLECTOR (wash review), VISION (ANPR) and TRAINER (phase-B job) container images and push them to the
 # house Gitea registry, tagged by BRANCH + short SHA (branch-aware: dev→:dev, stage→:stage,
 # main→:main). Separate from ci.yml (checks-only) and release.yml (tag-only desktop bundle).
 # Mirrors the house pattern (cf. trm/processor build.yml). See
@@ -14,6 +14,7 @@ on:
       - 'apps/web/**'
       - 'apps/vision/**'
       - 'apps/collector/**'
+      - 'apps/trainer/**'
       - 'packages/**'
       - 'package.json'
       - 'pnpm-lock.yaml'
@@ -61,6 +62,11 @@ jobs:
         working-directory: apps/vision
         run: uv sync --frozen
 
+      - name: Sync trainer deps
+        # Light core only — NOT the `train` extra (CPU torch, ~200 MB); the torch tests skip.
+        working-directory: apps/trainer
+        run: uv sync --frozen
+
       # Don't publish a broken image — run the same checks as ci.yml first.
       - name: Build + lint + test (Turbo)
         run: pnpm turbo run build lint test
@@ -119,12 +125,29 @@ jobs:
           context: apps/vision
           file: apps/vision/Dockerfile
           push: true
+          # The phase-B body-type classifier is fetched from the Gitea generic package registry
+          # at build when apps/vision/models/bodytype.version pins a version (empty = none). The
+          # registry user's credentials double as the fetch auth (BuildKit secret, never a layer).
+          secrets: |
+            bodytype_auth=${{ secrets.REGISTRY_USERNAME }}:${{ secrets.REGISTRY_PASSWORD }}
           tags: |
             ${{ env.REGISTRY }}/parking-vision:${{ steps.meta.outputs.branch }}
             ${{ env.REGISTRY }}/parking-vision:${{ steps.meta.outputs.branch }}-${{ steps.meta.outputs.sha }}
           cache-from: type=registry,ref=${{ env.REGISTRY }}/parking-vision:buildcache
           cache-to: type=registry,ref=${{ env.REGISTRY }}/parking-vision:buildcache,mode=max
 
+      - name: Build & push TRAINER (phase-B job)
+        uses: docker/build-push-action@v5
+        with:
+          context: apps/trainer
+          file: apps/trainer/Dockerfile
+          push: true
+          tags: |
+            ${{ env.REGISTRY }}/parking-trainer:${{ steps.meta.outputs.branch }}
+            ${{ env.REGISTRY }}/parking-trainer:${{ steps.meta.outputs.branch }}-${{ steps.meta.outputs.sha }}
+          cache-from: type=registry,ref=${{ env.REGISTRY }}/parking-trainer:buildcache
+          cache-to: type=registry,ref=${{ env.REGISTRY }}/parking-trainer:buildcache,mode=max
+
       # Optional: trigger a Komodo stack redeploy (cf. trm/processor). Enable by setting the
       # KOMODO_* secrets; left guarded so it no-ops until the parking stack is wired.
       - name: Trigger Komodo redeploy

.gitea/workflows/ci.yml

@@ -48,6 +48,11 @@ jobs:
         working-directory: apps/vision
         run: uv sync --frozen
 
+      - name: Sync trainer deps
+        # Same rule: light core only, not the `train` extra (CPU torch); torch tests skip.
+        working-directory: apps/trainer
+        run: uv sync --frozen
+
       - name: Build + lint (Turbo)
         # Covers tsc typecheck, vite build, i18n catalog type-parity (a missing sq/en
         # key fails the build), AND the vision service's ruff lint via uv.