feat: dataset export endpoint + SQLite dev-mode (geometry decoupling)

Dataset export (source data for model training):
- GET /api/v1/export?format=jsonl|csv|geojson with category/data_source
  filters; streams a labeled dataset (signal params + identified device +
  routed category) suitable for training a Sub-GHz classifier.

SQLite dev-mode (corrects FABLE brief: SQLite was NOT a drop-in swap):
- models.py made dialect-aware — JSONB->JSON, ARRAY(Text)->JSON, TSVECTOR
  ->Text via .with_variant(); PostGIS Geometry column + GiST index only
  defined when not on SQLite (lat/lon + haversine bbox used instead).
- config/database.py honors DATABASE_URL / a full-URL override and builds
  a SQLite engine (check_same_thread=False, no server pool) when the URL
  is sqlite; PostgreSQL keeps pooling + UTC session.

Verified: create_all + Capture/CaptureMatch/Device CRUD + JSON round-trip
+ bbox query all work on sqlite; postgres mode still defines geom + gist
index; 52/52 unit tests pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/api/main_simple.py

@@ -4,15 +4,17 @@ GigLez FastAPI Application - Simplified Version
 Runs without database requirement for testing web interface
 """
 
+import csv
+import io
 import json
 import sys
 from pathlib import Path
-from typing import List
+from typing import List, Optional
 from datetime import datetime
 from fastapi import FastAPI, Request, UploadFile, File, Form
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.middleware.gzip import GZipMiddleware
-from fastapi.responses import HTMLResponse
+from fastapi.responses import HTMLResponse, StreamingResponse
 from fastapi.staticfiles import StaticFiles
 from fastapi.templating import Jinja2Templates
 
@@ -240,6 +242,94 @@ async def get_stats():
     }
 
 
+# =============================================================================
+# DATASET EXPORT (for model training)
+# =============================================================================
+
+# Columns exported for training datasets. Order is stable across formats.
+EXPORT_FIELDS = [
+    "id", "filename", "frequency", "protocol", "preset",
+    "device_name", "device_category", "match_confidence", "match_method",
+    "latitude", "longitude", "gps_source", "timestamp",
+    "data_source", "session_id",
+]
+
+
+def _iter_export_rows(category: Optional[str], data_source: Optional[str]):
+    """Yield capture dicts filtered by category / data_source."""
+    for cap in captures_storage:
+        if category and cap.get("device_category") != category:
+            continue
+        if data_source and cap.get("data_source") != data_source:
+            continue
+        yield cap
+
+
+@app.get("/api/v1/export")
+async def export_dataset(
+    format: str = "jsonl",
+    category: Optional[str] = None,
+    data_source: Optional[str] = None,
+):
+    """
+    Export captures as a labeled dataset for model training.
+
+    Query params:
+    - format:      jsonl (default) | csv | geojson
+    - category:    filter to one device category (e.g. "Weather Sensor")
+    - data_source: filter to one data source (e.g. "production")
+
+    Each row carries the signal parameters + the auto-identified device
+    label + category, suitable as training data for a Sub-GHz classifier.
+
+    Note: this JSON-backed store does not retain raw pulse timing arrays;
+    for RAW-pulse feature training, use the DB-backed path (Capture.raw_data).
+    """
+    fmt = format.lower()
+    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+
+    if fmt == "jsonl":
+        def gen():
+            for cap in _iter_export_rows(category, data_source):
+                yield json.dumps({k: cap.get(k) for k in EXPORT_FIELDS}) + "\n"
+        return StreamingResponse(
+            gen(),
+            media_type="application/x-ndjson",
+            headers={"Content-Disposition": f'attachment; filename="giglez_dataset_{ts}.jsonl"'},
+        )
+
+    if fmt == "csv":
+        def gen():
+            buf = io.StringIO()
+            writer = csv.DictWriter(buf, fieldnames=EXPORT_FIELDS, extrasaction="ignore")
+            writer.writeheader()
+            yield buf.getvalue()
+            for cap in _iter_export_rows(category, data_source):
+                buf.seek(0); buf.truncate(0)
+                writer.writerow({k: cap.get(k) for k in EXPORT_FIELDS})
+                yield buf.getvalue()
+        return StreamingResponse(
+            gen(),
+            media_type="text/csv",
+            headers={"Content-Disposition": f'attachment; filename="giglez_dataset_{ts}.csv"'},
+        )
+
+    if fmt == "geojson":
+        features = []
+        for cap in _iter_export_rows(category, data_source):
+            lat, lon = cap.get("latitude"), cap.get("longitude")
+            if lat is None or lon is None:
+                continue
+            features.append({
+                "type": "Feature",
+                "geometry": {"type": "Point", "coordinates": [lon, lat]},
+                "properties": {k: cap.get(k) for k in EXPORT_FIELDS if k not in ("latitude", "longitude")},
+            })
+        return {"type": "FeatureCollection", "features": features}
+
+    return {"error": f"Unsupported format '{format}'. Use jsonl, csv, or geojson."}
+
+
 @app.delete("/api/v1/admin/cleanup")
 async def cleanup_test_data(data_source: str = "test"):
     """