surface the new analysis caveats in the workbench

The backend now says when a result covers a different area than was drawn, or
when a score belongs to one confidence cut only. None of that helps an
operator while it stays in the response body.

- the raster selection adapters carry the model-coverage and widened-cell
  warnings into the map panel and mark the result an estimate when either
  applies, so an existing warning slot renders them;
- the map workspace shows the selection-edge disclosure next to the object
  count;
- the detection panel shows average precision and the F1-optimal threshold
  beside the single-threshold figures, and the box-versus-footprint
  interpretation when candidates are detector boxes.

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

frontend/src/types.ts

@@ -505,6 +505,8 @@ export interface TerrainSelectionResponse {
   sample_count: number
   slope_sample_count: number
   coverage_ratio: number
+  /** Set when the selection was finer than one source cell and widened. */
+  cell_selection_warning?: string | null
   resolution_m: number
   vertical_reference: string
   summary: {
@@ -558,8 +560,14 @@ export interface FloodHazardSelectionResponse {
   selection_bbox: VectorSelectionBBox
   selection_area_id?: string | null
   selected_cell_count: number
+  /** Cells the flood model actually covers inside the selection. */
+  valid_cell_count?: number
+  no_data_cell_count?: number
+  data_coverage_ratio?: number
   inundated_cell_count: number
-  inundated_fraction: number
+  /** Share of the *modelled* cells; null when nothing was modelled. */
+  inundated_fraction: number | null
+  coverage_warning?: string | null
   resolution_m: number
   summary: {
     metric_label: string
@@ -633,6 +641,8 @@ export interface BathymetryRasterSelectionResponse {
   selected_cell_count: number
   valid_cell_count: number
   coverage_ratio: number
+  /** Set when the selection was finer than one source cell and widened. */
+  cell_selection_warning?: string | null
   resolution_m: number
   vertical_reference: 'mDNG'
   survey_period: string
@@ -690,6 +700,8 @@ export interface ThematicRasterSelectionResponse {
   selected_cell_count: number
   valid_cell_count: number
   coverage_ratio: number
+  /** Set when the selection was finer than one source cell and widened. */
+  cell_selection_warning?: string | null
   resolution_m: number
   observation_year: number
   summary: {
@@ -749,7 +761,14 @@ export interface VectorSelectionSummary {
   metric_unit: string
   aggregation_method: string
   primary_metric_key?: string | null
+  /**
+   * Whole features touching the selection. Area and length metrics clip to the
+   * selection, so these two fields say how far the populations diverge.
+   */
   feature_count: number
+  fully_covered_feature_count?: number | null
+  partially_covered_feature_count?: number | null
+  selection_edge_warning?: string | null
   is_estimate: boolean
   warning?: string | null
   metrics?: VectorSelectionMetric[]
@@ -1435,6 +1454,9 @@ export interface DetectionQaResult {
     canonical_method: string
     diagnostic_method: string
     iou_threshold: number
+    /** Whether the candidates are detector boxes or true footprint polygons. */
+    candidate_geometry_mode?: string
+    interpretation?: string
     strict_matches: number
     envelope_matches: number
     possible_box_to_footprint_mismatch_count: number
@@ -1444,7 +1466,37 @@ export interface DetectionQaResult {
     envelope_recall?: number | null
     envelope_f1_score?: number | null
     envelope_mean_iou?: number | null
+    envelope_precision_recall_curve?: PrecisionRecallCurve
   }
+  precision_recall_curve?: PrecisionRecallCurve
+}
+
+/**
+ * Threshold-independent view of a detection run: precision and recall at every
+ * confidence value that occurs, so two models can be compared without both
+ * being read at one arbitrary cut.
+ */
+export interface PrecisionRecallCurve {
+  iou_threshold: number
+  reference_count: number
+  candidate_count: number
+  average_precision: number
+  best_f1: number
+  best_f1_threshold: number | null
+  best_f1_precision?: number | null
+  best_f1_recall?: number | null
+  points: PrecisionRecallPoint[]
+}
+
+export interface PrecisionRecallPoint {
+  confidence_threshold: number
+  candidate_count: number
+  true_positives: number
+  false_positives: number
+  false_negatives: number
+  precision: number
+  recall: number
+  f1_score: number
 }
 
 export type SegmentationModelCapability = DetectionModelCapability