document the analysis contracts and correct the job-queue claim

The README's architecture diagram and stack list promised Redis + RQ. There is
no such thing in this codebase: redis_url is a settings field nothing reads,
rq is not a dependency, and the only worker is an asyncio poller over a
Postgres job table. Describe what exists.

API_CONTRACTS and KNOWN_LIMITATIONS gain the rules a reader needs to interpret
a result: deterministic confidence-ranked matching, the precision/recall curve,
the three flood-hazard cell populations, the sub-cell selection fallback, the
whole-feature object count next to clipped area metrics, and the required
manifest CRS. The new limitations are stated as limits rather than buried:
raster analysis has no sub-cell precision, and the object count is not
fractional.

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

.env.example

@@ -131,7 +131,13 @@ YOLO_IMAGE_SIZE=640
 YOLO_MAX_TILES=100
 YOLO_MAX_DETECTIONS=1000
 YOLO_DUPLICATE_IOU_THRESHOLD=0.5
+# Tiles per model.predict call. 1 leaves most of a modern GPU idle on a run of
+# a hundred tiles; raise it until VRAM is the limit.
 YOLO_BATCH_SIZE=1
+# Drop boxes truncated by an interior tile edge. Because tiles overlap, the
+# neighbouring tile saw the same object whole, so the truncated half is a
+# duplicate and a shape error at once. Boxes on the outer raster edge are kept.
+YOLO_SUPPRESS_TILE_EDGE_DETECTIONS=true
 
 # Local segmentation models. GeoIntel never downloads model weights
 # automatically; point these to existing local files to enable inference.
@@ -176,3 +182,7 @@ GEOINTEL_CORS_ORIGINS=https://geointel.itworx.tech,http://geointel.itworx.tech,h
 GEOINTEL_MAX_UPLOAD_MB=500
 GEOINTEL_AOI_WORKER_ENABLED=false
 GEOINTEL_AOI_WORKER_POLL_SECONDS=2
+# Executes queued detection.run / segmentation.run jobs from POST
+# /detection/run-async, so tiled GPU inference never blocks an HTTP request.
+GEOINTEL_ANALYSIS_WORKER_ENABLED=false
+GEOINTEL_ANALYSIS_WORKER_POLL_SECONDS=2