feat: scope detection QA to inference coverage

docs/AI_PIPELINES.md

@@ -544,6 +544,15 @@ Sprint 8C makes persisted detections reviewable:
 - Persisted detection geometries can be returned as GeoJSON FeatureCollections for MapLibre display.
 - Detection QA compares candidate detection geometries against persisted reference `vector_features`.
 - QA results reuse `quality_checks` and `metrics`; no parallel QA persistence system is introduced.
+- Configured-YOLO QA derives its evaluation extent from the persisted tile
+  manifest. Tile bounds are transformed from their explicit source CRS to
+  EPSG:4326, unioned, and used to clip candidate/reference populations before
+  canonical footprint-IoU matching. Reference features wholly outside the
+  imagery presented to the model no longer count as false negatives.
+- A separate reference-envelope IoU pass is persisted as
+  `box_to_footprint_diagnostics`. It quantifies possible matching artifacts from
+  comparing rectangular detections with irregular building footprints, but is
+  diagnostic only and never changes canonical QA metrics.
 - Segmentation remains out of scope for Sprint 8C.
 
 ## 2. Tile Metadata

docs/API_CONTRACTS.md

@@ -872,6 +872,29 @@ Response persists a `quality_check` and `metrics` rows through the existing QA/Q
 - `false_negatives`
 - `quality_check_id`
 
+Configured-YOLO QA automatically reads `tile_manifest_path` from the persisted
+`AnalysisRun.parameters_json`. Candidate and reference geometries are clipped
+to the union of the manifest's tile bounds after explicit CRS transformation to
+EPSG:4326. The response additionally returns:
+
+- `candidate_feature_count_raw` and `reference_feature_count_raw`;
+- `coverage`, including raw/evaluated/excluded/boundary-clipped population
+  counts, tile count, source CRS values and coverage mode;
+- `box_to_footprint_diagnostics`, which compares candidate boxes with reference
+  envelopes at the same IoU threshold.
+
+The canonical precision, recall, F1 and mean IoU always remain based on
+candidate geometry versus the persisted reference footprint. Envelope results
+are explicitly `diagnostic_only` and are persisted in
+`quality_checks.findings_json`; they never replace or inflate canonical metrics.
+
+Configured-YOLO QA fails closed with `DETECTION_QA_COVERAGE_UNAVAILABLE` when
+manifest provenance is absent, `DETECTION_QA_COVERAGE_MISMATCH` when it belongs
+to another raster, `DETECTION_QA_COVERAGE_INVALID` when bounds/CRS are invalid,
+or `REFERENCE_FEATURES_OUTSIDE_COVERAGE` when no reference polygons overlap the
+actual inference coverage. Explicit fixture/legacy runs without a manifest keep
+the documented unbounded comparison behavior.
+
 If the reference dataset has no persisted vector features, the endpoint returns `REFERENCE_FEATURES_NOT_FOUND`. It does not calculate fake QA metrics.
 
 #### Future analysis route: `/api/v1/analysis/building-stats`

docs/CODEX_EXECUTION_LOG.md

@@ -7556,3 +7556,46 @@ Open:
 - Clip detection QA reference populations to actual raster/tile coverage and
   add box-to-building-footprint matching diagnostics before any further model
   training or promotion decision.
+
+# Sprint 184 - Detection QA coverage and matching diagnostics
+
+## Implementation
+
+- Added a focused detection-QA coverage service that resolves the persisted
+  tile manifest from the analysis run, validates dataset ownership and CRS,
+  transforms each tile footprint to EPSG:4326 and unions the exact inference
+  coverage.
+- Configured-YOLO QA now fails closed when persisted manifest provenance is
+  absent, invalid or belongs to another raster. Explicit fixture and legacy
+  runs without a manifest retain their existing unbounded QA behavior.
+- Candidate Detection geometries and persisted reference VectorFeature
+  geometries are clipped to inference coverage before the existing one-to-one
+  IoU matcher runs. Raw, evaluated, excluded-outside and boundary-clipped
+  counts are persisted in `quality_checks.findings_json`.
+- Canonical precision, recall, F1, mean IoU and metric rows remain strict
+  candidate-polygon versus reference-footprint results. A second
+  candidate-polygon versus reference-envelope pass is persisted and displayed
+  as diagnostic-only evidence; it never replaces or promotes canonical
+  metrics.
+- Detection Lab now explains which reference population was evaluated and
+  clearly separates possible box-to-footprint artifacts from the canonical
+  scorecard.
+- Updated the real-data operator smoke assertions, API/AI/database contracts
+  and backend/frontend operator documentation. No migration, request contract,
+  provider fetch or model dependency changed.
+
+## Local validation
+
+- `bash scripts/run_readiness_check.sh` passed end to end.
+- Full backend suite passed: `506` tests.
+- API contract audit passed with `81` implemented routes and the two documented
+  non-envelope endpoints.
+- Alembic reports one head: `202606120900`.
+- Frontend TypeScript checking and production build passed with `85` modules;
+  the dedicated MapLibre chunk remains intact.
+
+## Next pass
+
+- Deploy Sprint 184 to Tower, rerun QA for the persisted Mol-center analysis
+  run and verify both persisted coverage evidence and Detection Lab rendering
+  against live PostGIS before making any model-training decision.

docs/DATABASE_IMPLEMENTATION_PLAN.md

@@ -189,6 +189,13 @@ Metrics may belong to a quality check, an analysis run, or both. Sprint 7A persi
 
 Quality checks are domain records. Jobs track execution state; quality checks track the persisted QA/QC result; metrics track individual measurements.
 
+Detection QA coverage and box-to-footprint diagnostics require no schema
+change. The canonical metric rows remain precision, recall, F1, mean IoU and
+false-positive/negative counts. Tile coverage population counts and the
+diagnostic reference-envelope comparison are persisted in the existing
+`quality_checks.findings_json`; `parameters_json.coverage_policy` records the
+evaluation policy used for reproducibility.
+
 ### exports
 
 - `id uuid primary key`

docs/TODO.md

@@ -26,7 +26,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Separate persisted database layers from available analysis-result overlays with an explicit Map content mode.
 - [x] Persist combined Mol operator evidence under the Unraid storage mount so reports survive all-in-one container replacement.
 - [x] Visually review Mol Postel and Donk false-positive/false-negative evidence, classify the dominant error modes and only then decide whether another model-training pass is justified.
-- [ ] Clip detection QA populations to persisted raster/tile coverage and add box-to-footprint matching diagnostics before reconsidering model training.
+- [x] Clip detection QA populations to persisted raster/tile coverage and add box-to-footprint matching diagnostics before reconsidering model training.
 - [x] Backend FastAPI foundation, health endpoint and service structure.
 - [x] React/TypeScript frontend foundation and MapLibre workbench.
 - [x] Map layer visibility, opacity and feature property inspection.
@@ -411,7 +411,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Bound populated Project, Dataset, QA, AI and Export panels so long histories do not push core actions thousands of pixels down-page.
 - [x] Move Data create/upload forms and Map/AI diagnostics into explicit progressive disclosures.
 - [x] Reorder AI Labs around run controls and Map around layer selection plus the MapLibre canvas.
-- [ ] Harden detection QA coverage and matching diagnostics before any further model training.
+- [x] Harden detection QA coverage and matching diagnostics before any further model training.
 
 - [x] Replace the one-page workflow panel stack with a task-based workbench shell.
 - [x] Add persistent project/AOI/dataset/layer context.