feat: add measured detection review loop

frontend/README.md

@@ -8,6 +8,12 @@ The user-facing shell is task based: `Kaart`, `Bronnen`, `Kwaliteit`, `Beeldanal
 
 Detection defaults to the configured local YOLO asset and automatically selects an available raster and active model asset where possible. The model registry and preflight remain honest when PyTorch, Ultralytics, a local model file or a tile manifest is unavailable. The active building profile is operational but remains review-required: its recorded coverage-aware benchmark is approximately precision 0.590, recall 0.577 and F1 0.582 over seven positive AOIs, with zero detections in the empty-background control. Another training pass is intentionally blocked until the generated false-positive and false-negative review decisions are completed.
 
+Map-driven building analysis uses the documented footprint-IoU `0.25` and
+distinguishes model candidates from verified buildings. It shows persisted
+matches, precision, recall, F1, false positives and false negatives. The
+Quality workspace includes a paginated review queue for persisted detection-QA
+evidence. Reviews do not rewrite detections, GRB geometry or QA metrics.
+
 The map-first explorer has two deliberate modes. `Latest state` selects the
 latest explicitly dated source snapshot without claiming an old edition is
 current, while `Evolution` lets the operator compare an earlier and later
@@ -373,6 +379,7 @@ Before creating tiles, the guided action inspects raster dimensions and estimate
 - The QA/QC workspace includes a selected-check evidence drilldown with candidate/reference provenance, false-positive/negative metric evidence, map handoff context and parameters/findings JSON.
 - QA/QC findings now persist feature-level evidence in `findings_json`: matched candidate/reference feature ids with IoU, false-positive candidate feature ids and false-negative reference feature ids. The QA/QC drilldown renders these as compact evidence lists before the raw JSON.
 - Persisted QA/QC checks can be rendered as a Map workspace evidence overlay. The QA/QC panel calls `GET /api/v1/projects/{project_id}/quality-checks/{quality_check_id}/evidence/geojson`, then MapLibre draws matched candidate/reference geometries, false positives and false negatives with distinct styling and a compact legend.
+- Detection QA checks expose a filtered, paginated operator review queue through `GET/POST /api/v1/projects/{project_id}/quality-checks/{quality_check_id}/reviews`. The UI keeps confirmed model errors separate from reference gaps and box-to-footprint alignment mismatches.
 
 ## Raster dependency visibility