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>

config/database.py

@@ -24,7 +24,8 @@ class DatabaseConfig:
         password: str = "giglez_secure_password_2026",
         pool_size: int = 10,
         max_overflow: int = 20,
-        echo: bool = False
+        echo: bool = False,
+        url: Optional[str] = None
     ):
         """
         Initialize database configuration
@@ -38,6 +39,9 @@ class DatabaseConfig:
             pool_size: Connection pool size (Wigle pattern: moderate pooling)
             max_overflow: Max overflow connections
             echo: Echo SQL queries (debug mode)
+            url: Full SQLAlchemy URL override (e.g. sqlite:///./giglez.db).
+                 When set, it takes precedence over the host/port/user fields.
+                 Falls back to the DATABASE_URL env var.
         """
         self.host = host
         self.port = port
@@ -47,29 +51,50 @@ class DatabaseConfig:
         self.pool_size = pool_size
         self.max_overflow = max_overflow
         self.echo = echo
+        self.url = url or os.getenv("DATABASE_URL")
 
         self._engine: Optional = None
         self._session_factory: Optional[sessionmaker] = None
 
+    @property
+    def is_sqlite(self) -> bool:
+        return bool(self.url) and self.url.startswith("sqlite")
+
     @property
     def connection_string(self) -> str:
-        """Generate PostgreSQL connection string"""
+        """SQLAlchemy connection string (URL override wins, else PostgreSQL)"""
+        if self.url:
+            return self.url
         return f"postgresql://{self.user}:{self.password}@{self.host}:{self.port}/{self.database}"
 
     @property
     def connection_string_safe(self) -> str:
-        """Generate connection string without password (for logging)"""
+        """Connection string without password (for logging)"""
+        if self.url:
+            # sqlite URLs carry no password; postgres URLs would — mask them
+            if self.is_sqlite:
+                return self.url
+            return "postgresql://****"
         return f"postgresql://{self.user}:****@{self.host}:{self.port}/{self.database}"
 
     def get_engine(self):
         """
-        Get or create SQLAlchemy engine
+        Get or create SQLAlchemy engine.
 
-        Uses connection pooling for performance (Wigle pattern)
+        PostgreSQL uses connection pooling + UTC session (Wigle pattern).
+        SQLite (dev/MVP) uses a simple engine with cross-thread access enabled.
         """
         if self._engine is None:
             logger.info(f"Creating database engine: {self.connection_string_safe}")
 
+            if self.is_sqlite:
+                # SQLite dev mode: no server-side pool, allow use across threads
+                self._engine = create_engine(
+                    self.connection_string,
+                    echo=self.echo,
+                    connect_args={"check_same_thread": False},
+                )
+            else:
                 self._engine = create_engine(
                     self.connection_string,
                     poolclass=QueuePool,

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"):
     """

src/database/models.py

@@ -5,17 +5,38 @@ Database models matching the PostgreSQL + PostGIS schema
 Based on Wigle wardriving patterns adapted for IoT RF device mapping
 """
 
+import os
 from datetime import datetime
 from typing import Optional, List
 from sqlalchemy import (
     Column, String, Integer, Float, DateTime, Text, Boolean,
     DECIMAL, ARRAY, ForeignKey, CheckConstraint, UniqueConstraint,
-    Index, LargeBinary
+    Index, LargeBinary, JSON
 )
 from sqlalchemy.orm import declarative_base, relationship
-from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy.dialects.postgresql import JSONB, TSVECTOR
-from geoalchemy2 import Geometry
+
-from geoalchemy2.functions import ST_SetSRID, ST_MakePoint
+# =============================================================================
+# DIALECT AWARENESS (PostgreSQL/PostGIS in prod, SQLite in dev)
+# =============================================================================
+# The production schema targets PostgreSQL + PostGIS, but dev/MVP must run on
+# SQLite with no external services. We keep PostgreSQL-native behaviour on
+# postgres via `.with_variant()` while degrading gracefully on SQLite.
+
+_DATABASE_URL = os.getenv("DATABASE_URL", "")
+IS_SQLITE = _DATABASE_URL.startswith("sqlite")
+
+# JSON: JSONB on postgres, generic JSON (stored as TEXT) on sqlite
+JSON_TYPE = JSON().with_variant(JSONB(), "postgresql")
+# Text arrays: native ARRAY on postgres, JSON list on sqlite
+STR_ARRAY_TYPE = JSON().with_variant(ARRAY(Text), "postgresql")
+# Full-text search vector: TSVECTOR on postgres (trigger-populated), Text on sqlite
+SEARCH_VECTOR_TYPE = Text().with_variant(TSVECTOR(), "postgresql")
+
+if not IS_SQLITE:
+    # geoalchemy2 is only importable/usable against a spatial backend
+    from geoalchemy2 import Geometry
+    from geoalchemy2.functions import ST_SetSRID, ST_MakePoint
 
 
 # =============================================================================
@@ -154,7 +175,7 @@ class Device(Base):
     typical_frequency = Column(Integer, index=True)
     frequency_range_low = Column(Integer)
     frequency_range_high = Column(Integer)
-    modulation_types = Column(ARRAY(Text))
+    modulation_types = Column(STR_ARRAY_TYPE)
 
     # Protocol Information
     protocol = Column(String(100), index=True)
@@ -174,8 +195,8 @@ class Device(Base):
     source = Column(String(50), index=True)  # 'flipper', 'rtl433', 'urh', 'community'
     source_url = Column(Text)
 
-    # Full-text search (auto-updated by trigger)
+    # Full-text search (auto-updated by trigger on postgres)
-    search_vector = Column('search_vector', nullable=True)
+    search_vector = Column('search_vector', SEARCH_VECTOR_TYPE, nullable=True)
 
     # Relationships
     signatures = relationship("Signature", back_populates="device")
@@ -218,6 +239,8 @@ class Capture(Base):
     longitude = Column(DECIMAL(11, 8), nullable=False)
     altitude = Column(DECIMAL(8, 2))
     gps_accuracy = Column(DECIMAL(6, 2))
+    # Spatial geometry only exists on PostGIS; SQLite uses lat/lon + haversine
+    if not IS_SQLITE:
         geom = Column(Geometry('POINT', srid=4326))  # Auto-populated by trigger
 
     # Timestamps
@@ -254,7 +277,10 @@ class Capture(Base):
         CheckConstraint('longitude >= -180 AND longitude <= 180', name='valid_longitude'),
         CheckConstraint('match_confidence >= 0 AND match_confidence <= 1', name='valid_confidence'),
         CheckConstraint('frequency >= 300000000 AND frequency <= 928000000', name='valid_frequency'),
-        Index('idx_captures_geom', 'geom', postgresql_using='gist'),
+    ) + (
+        # GiST spatial index only applies to the PostGIS geometry column
+        (Index('idx_captures_geom', 'geom', postgresql_using='gist'),)
+        if not IS_SQLITE else ()
     )
 
     # Relationships
@@ -350,7 +376,7 @@ class CaptureMatch(Base):
     # Match Details
     confidence = Column(DECIMAL(5, 4), nullable=False, index=True)
     match_method = Column(String(50), nullable=False)
-    match_details = Column(JSONB)
+    match_details = Column(JSON_TYPE)
 
     # Timestamp
     matched_at = Column(DateTime, default=datetime.utcnow)
@@ -386,7 +412,7 @@ class Identification(Base):
     notes = Column(Text)
 
     # Visual Evidence
-    photo_urls = Column(ARRAY(Text))
+    photo_urls = Column(STR_ARRAY_TYPE)
 
     # Community Validation
     upvotes = Column(Integer, default=0)
@@ -540,7 +566,7 @@ class RTL433Protocol(Base):
     bit_count = Column(Integer)
 
     # JSON Fields Mapping
-    json_fields = Column(JSONB)
+    json_fields = Column(JSON_TYPE)
 
     # Source
     source_file = Column(String(500))