feat(ocr): add metrics.py factory with test-scoped CollectorRegistry support

Encapsulates every custom OCR metric in an OcrMetrics frozen dataclass and
exposes a `build_metrics(registry)` factory. Production main.py binds against
the default REGISTRY; tests construct a fresh CollectorRegistry per case and
monkeypatch main.metrics, so counter values stay isolated between tests
(decision #3 on issue #652, Option A).

Refs #652

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ocr-service/metrics.py

@@ -0,0 +1,92 @@
+"""Prometheus metric definitions for the OCR service.
+
+`build_metrics(registry)` returns a fresh `OcrMetrics` instance bound to the
+given `CollectorRegistry`. Production code calls it once at module load with
+the default `REGISTRY`; tests pass a per-test `CollectorRegistry()` to keep
+counter values isolated between cases (decision #3 on issue #652).
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from prometheus_client import CollectorRegistry, Counter, Gauge, Histogram
+
+
+@dataclass(frozen=True)
+class OcrMetrics:
+    """Container for every custom OCR metric.
+
+    Counters and gauges are immutable references to `prometheus_client`
+    instances. Mutating them (`.inc()`, `.observe()`, `.set()`) is safe;
+    rebinding the field on the dataclass is not — use `build_metrics` to get
+    a new container.
+    """
+
+    ocr_jobs_total: Counter
+    ocr_pages_total: Counter
+    ocr_skipped_pages_total: Counter
+    ocr_words_total: Counter
+    ocr_illegible_words_total: Counter
+    ocr_processing_seconds: Histogram
+    ocr_training_runs_total: Counter
+    ocr_model_accuracy: Gauge
+    ocr_models_ready: Gauge
+
+
+def build_metrics(registry: CollectorRegistry) -> OcrMetrics:
+    """Create one OcrMetrics instance bound to `registry`."""
+    return OcrMetrics(
+        ocr_jobs_total=Counter(
+            "ocr_jobs_total",
+            "Number of OCR jobs processed, labelled by engine and script type.",
+            ["engine", "script_type"],
+            registry=registry,
+        ),
+        ocr_pages_total=Counter(
+            "ocr_pages_total",
+            "Number of pages successfully OCR'd, labelled by engine.",
+            ["engine"],
+            registry=registry,
+        ),
+        ocr_skipped_pages_total=Counter(
+            "ocr_skipped_pages_total",
+            "Number of pages skipped because the OCR engine raised.",
+            registry=registry,
+        ),
+        ocr_words_total=Counter(
+            "ocr_words_total",
+            "Number of words recognized across all OCR blocks.",
+            registry=registry,
+        ),
+        ocr_illegible_words_total=Counter(
+            "ocr_illegible_words_total",
+            "Number of words below the confidence threshold "
+            "(replaced with [unleserlich]).",
+            registry=registry,
+        ),
+        ocr_processing_seconds=Histogram(
+            "ocr_processing_seconds",
+            "OCR processing time per page (streaming) or per document (non-streaming).",
+            ["engine"],
+            registry=registry,
+        ),
+        ocr_training_runs_total=Counter(
+            "ocr_training_runs_total",
+            "Number of training runs, labelled by kind (recognition|segmentation) "
+            "and outcome (success|error).",
+            ["kind", "outcome"],
+            registry=registry,
+        ),
+        ocr_model_accuracy=Gauge(
+            "ocr_model_accuracy",
+            "Latest model accuracy reported by a successful training run.",
+            ["kind"],
+            registry=registry,
+        ),
+        ocr_models_ready=Gauge(
+            "ocr_models_ready",
+            "1 once the lifespan startup has finished loading models, 0 before.",
+            registry=registry,
+        ),
+    )