Add persisted false-positive visual review gate

docs/AI_PIPELINES.md

@@ -129,6 +129,27 @@ candidates with EPSG:4326 geometry IoU greater than or equal to
 raw, persisted and suppressed detection counts so calibration evidence remains
 auditable.
 
+### Persisted false-positive visual review
+
+Detection QA labels a candidate as a false-positive only relative to the
+selected persisted reference dataset and matching tolerance. That finding is
+not automatically a model error: the reference can be incomplete or stale, and
+alignment can be wrong. GeoIntel therefore exposes persisted detection
+confidence/model/tile/bbox provenance in the existing QA evidence GeoJSON and
+provides a read-only contact-sheet workflow.
+
+The operator must explicitly select one of:
+
+- `confirmed_model_false_positive`;
+- `reference_gap_or_change`;
+- `qa_alignment_mismatch`;
+- `uncertain`;
+- `unreviewed`.
+
+Only records explicitly marked `confirmed_model_false_positive` are emitted by
+the validator as possible hard-negative review input. The workflow does not
+train a model, mutate QA persistence, fetch data or infer review decisions.
+
 ### Local model asset catalog
 
 GeoIntel can list local runtime model files mounted into the backend model

docs/API_CONTRACTS.md

@@ -1215,6 +1215,14 @@ Response:
 `false_positive` or `false_negative`. Missing persisted feature ids are reported
 in `warnings`; no fake geometries are produced.
 
+Detection-backed candidate evidence also exposes provenance read from the
+persisted `detections` row: `detection_id`, `job_id`, `confidence`,
+`model_name`, `model_version`, `source_tile_path` and `bbox_json`. Existing
+`properties_json` fields such as `tile_index` remain present. Segmentation-backed
+candidate evidence exposes the equivalent persisted model/source fields plus
+`segmentation_id`, `mask_path` and `area_m2`. These are additive GeoJSON
+properties; the canonical envelope and endpoint path are unchanged.
+
 ## Exports
 
 ### POST `/api/v1/exports/geojson`

docs/CODEX_EXECUTION_LOG.md

@@ -7189,3 +7189,38 @@ Open:
 ## Next recommended pass
 
 - Visually classify a stratified false-positive sample from Turnhout, Herentals and Geel before deciding whether any confirmed examples belong in a new hard-negative corpus. Review the remaining 5,838 persistent false negatives in the same evidence-led pass; do not start another blind training run.
+
+# Sprint 176 - Detection false-positive visual review gate
+
+## Persisted provenance
+
+- Extended the existing read-only QA evidence GeoJSON conversion so detection-backed evidence carries the persisted detection id, job id, confidence, model name/version, source tile path and pixel bbox.
+- Added equivalent persisted segmentation provenance fields without changing the endpoint, canonical envelope, ORM or migration chain.
+- Historical `QualityCheck` evidence can be re-exported against existing persisted `Detection` rows; no QA rerun or data rewrite is required.
+
+## Manual visual review
+
+- Added a storage-root-confined contact-sheet renderer for persisted detection false-positive evidence.
+- The renderer validates portfolio role counts, polygon geometry, source imagery and persisted provenance, then selects deterministically across AOI, WGS84 area bucket and confidence band.
+- Source imagery is rendered with the candidate pixel bbox plus persisted matched-reference and missed-reference overlays.
+- Added an explicit five-state review CSV: `confirmed_model_false_positive`, `reference_gap_or_change`, `qa_alignment_mismatch`, `uncertain` and `unreviewed`.
+- Added a separate validator that rejects missing, duplicate, unexpected or invalid decisions. `--require-complete` exits with code `2` while any record remains unreviewed.
+- Only explicitly confirmed model false-positives are emitted to `confirmed_model_false_positives.geojson`; no QA result is automatically converted into a model label or training artifact.
+
+## Validation
+
+- `python -m compileall backend/app`: passed.
+- `python -m pytest`: 478 passed.
+- Focused provenance/render/path-confinement/incomplete-review/export tests: passed.
+- `python -m ruff check` for changed Python services, scripts and tests: passed.
+- Generated fixture contact sheet was visually inspected at 128 px thumbnails; candidate/reference/missed-reference overlays and header provenance remained readable.
+- `npm run typecheck`: passed.
+- `npm run build`: passed; app bundle `217.00 kB`, MapLibre bundle `801.82 kB` before gzip.
+- `bash scripts/run_readiness_check.sh`: passed with 478 tests and the new operator-script compile gates.
+- `python -m alembic heads`: one head, `202606120900`.
+- `python -m alembic upgrade head --sql`: complete migration chain rendered successfully.
+- Local Docker validation remains unavailable because Docker CLI is not installed on the Windows host; live all-in-one/PostGIS validation follows on Tower after deployment.
+
+## Next recommended pass
+
+- Re-export the seven-AOI evidence portfolio from the deployed backend, render the Turnhout/Herentals/Geel sheets and inspect the real orthophoto evidence. Keep all CSV decisions `unreviewed` until an operator makes an explicit visual classification; do not start another model training run yet.

docs/TODO.md

@@ -139,6 +139,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Complete rebuild/restart and browser/runtime smoke for the guarded promoted V1 building detector activation.
 - [x] Expand focused small-building training evidence after reviewing persistent false negatives, train one inactive candidate and pass it through positive, pure-empty and fixed-reference promotion evidence before guarded activation.
 - [x] Audit the promoted model's increased false-positive load from persisted seven-AOI QA evidence, including geodetic area buckets, AOI-qualified tile hotspots and combined review GeoJSON.
+- [x] Add persisted detection provenance, stratified visual contact sheets and an explicit manual-decision gate for false-positive review.
 - [ ] Visually classify representative false-positive evidence from Turnhout, Herentals and Geel and review the remaining 5,838 persistent false negatives before any further model training; do not start another blind run.
 
 ## Sprint 8 status