calibrate a confidence threshold from one inference pass

Threshold calibration ran the model over every tile once per threshold — three
GPU passes to compare 0.50, 0.25 and 0.15 on a hundred-tile raster. The answer
is already in a single run at the lowest value: detections above a higher cut
are a subset of it, and duplicate suppression walks candidates in descending
confidence, so a lower-confidence box can never displace a higher-confidence
one. The kept set above any cut is identical whichever threshold the run used,
which is what makes one pass sufficient rather than merely cheaper.

QA now takes calibration_thresholds and reads each operating point off the same
precision/recall walk it already performs, marking the F1-optimal cut. The lab
runs inference once and fills its table from the sweep.

The contract test asserted the per-threshold loop by name, pinning the waste it
was meant to describe. It now states what calibration owes an operator: a row
per requested threshold, from one run.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/app/services/detection_metrics_service.py

@@ -118,6 +118,66 @@ class DetectionMetricsService:
             area += (recalls[index] - recalls[index - 1]) * precisions[index]
         return area
 
+    @staticmethod
+    def operating_point(curve: dict[str, Any], *, min_confidence: float) -> dict[str, Any]:
+        """The metrics that hold when detections below ``min_confidence`` are dropped.
+
+        Read off the curve rather than recomputed: the curve already walked the
+        ranking once, and every threshold is a prefix of that walk. Running the
+        model again per threshold would spend N GPU passes to reproduce numbers
+        that are already here.
+        """
+
+        points = curve.get("points") or []
+        reference_count = int(curve.get("reference_count") or 0)
+        admitted = [point for point in points if point["confidence_threshold"] >= min_confidence]
+        # Points are cumulative down the ranking, so the last admitted one is
+        # the complete tally at this cut.
+        tally = admitted[-1] if admitted else None
+
+        true_positives = int(tally["true_positives"]) if tally else 0
+        false_positives = int(tally["false_positives"]) if tally else 0
+        false_negatives = max(0, reference_count - true_positives)
+        candidate_count = true_positives + false_positives
+
+        precision = true_positives / candidate_count if candidate_count else None
+        recall = true_positives / reference_count if reference_count else None
+        f1_score = None
+        if precision is not None and recall is not None:
+            f1_score = (2 * precision * recall / (precision + recall)) if precision + recall > 0 else 0.0
+
+        return {
+            "min_confidence": min_confidence,
+            "confidence_threshold": tally["confidence_threshold"] if tally else None,
+            "candidate_count": candidate_count,
+            "true_positives": true_positives,
+            "false_positives": false_positives,
+            "false_negatives": false_negatives,
+            "precision": precision,
+            "recall": recall,
+            "f1_score": f1_score,
+        }
+
+    @staticmethod
+    def calibration_sweep(curve: dict[str, Any], *, thresholds: list[float]) -> list[dict[str, Any]]:
+        """Every requested operating point, strictest first, from one curve.
+
+        Recall cannot fall as the cut loosens — that monotonicity is exactly
+        why a single run answers the whole sweep.
+        """
+
+        ordered = sorted({float(value) for value in thresholds}, reverse=True)
+        rows = [
+            DetectionMetricsService.operating_point(curve, min_confidence=value) for value in ordered
+        ]
+        best_f1 = max((row["f1_score"] or 0.0) for row in rows) if rows else 0.0
+        marked = False
+        for row in rows:
+            is_best = not marked and (row["f1_score"] or 0.0) == best_f1
+            row["best_f1_in_sweep"] = is_best
+            marked = marked or is_best
+        return rows
+
     @staticmethod
     def precision_recall_curve(
         candidates: list[tuple[dict[str, Any], BaseGeometry]],