docs: add README with map/identification screenshots and pipeline diagram

Hero screenshot of the live Leaflet device map plus a marker popup showing an
auto-identified device with confidence, and a Mermaid diagram of the
parse -> category-route -> match -> map/export pipeline. Documents simple-mode
run steps, API surface, and honest Phase-0 accuracy caveats.

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

README.md

@@ -0,0 +1,108 @@
+# GigLez — Wigle for Sub-GHz IoT
+
+**A crowdsourced map of Sub-GHz RF devices.** Upload a Flipper Zero `.sub` capture with GPS,
+GigLez parses the signal, identifies the device by protocol signature, and pins it on an
+interactive map — building an open dataset of the 300–928 MHz IoT world (garage doors,
+doorbells, weather sensors, TPMS, remotes, and more).
+
+![GigLez device map](docs/img/hero-map.png)
+
+Click any marker to see the automatically-identified device and match confidence:
+
+![Device identification popup](docs/img/device-popup.png)
+
+---
+
+## Why
+
+Wigle.net mapped the WiFi/Bluetooth/cellular world. Almost nobody has done the same for
+**Sub-GHz RF** — the unlicensed ISM bands where garage remotes, alarm sensors, tire-pressure
+monitors and cheap IoT gadgets chatter. GigLez collects those captures from *any* device
+(Flipper Zero, LilyGo, RTL-SDR, HackRF — anything that emits `.sub`), identifies them, and
+turns the aggregate into a labeled dataset for RF SIGINT research and ML training.
+
+## How it works
+
+```mermaid
+flowchart LR
+    A[".sub capture<br/>+ GPS"] --> B["Parser<br/>freq · preset · timing"]
+    B --> C["Category Router<br/>band + timing-ratio + pulse-count"]
+    C --> D["Protocol DB match<br/>+ confidence calibration"]
+    D --> E["GPS resolve<br/>filename / manifest"]
+    E --> F["Store"]
+    F --> G["Leaflet map"]
+    F --> H["Dataset export<br/>jsonl · csv · geojson"]
+```
+
+The identifier first **routes** a capture to a device category using frequency band, pulse
+timing ratios and pulse counts, then scores it only against protocols in that category and
+calibrates the confidence (spread + category-mismatch penalties). This routing step is what
+lifts identification accuracy over naive whole-database matching.
+
+## Features
+
+- **Device-agnostic upload** — drag-and-drop `.sub` files; GPS from the filename or a manifest
+- **Automatic identification** — category routing → protocol-signature matching → confidence score
+- **Interactive map** — Leaflet with marker clustering, frequency-band color coding, and a heatmap view
+- **Search & filter** — by frequency, protocol, and geographic radius/bounding box
+- **Dataset export** — stream the labeled corpus as JSONL, CSV, or GeoJSON for ML training
+- **Stats dashboard** — total captures, unique devices, geographic coverage
+
+## Quick start
+
+```bash
+# install deps into a venv
+python3 -m venv venv && source venv/bin/activate
+pip install -r requirements.txt
+
+# run the app (simple mode — JSON-backed, no database required)
+./start_web.sh
+# or explicitly:
+python3 -m uvicorn src.api.main_simple:app --host 0.0.0.0 --port 8000 --reload
+```
+
+Then open **http://localhost:8000** for the map UI (`/docs` for the OpenAPI reference).
+
+> **Modes.** `main_simple.py` is the live app and needs no database — captures persist to a
+> local JSON store, which is the fastest way to try GigLez. A full SQLAlchemy ORM path
+> (PostgreSQL + PostGIS in production, SQLite in dev via `DATABASE_URL=sqlite:///./giglez.db`)
+> also exists for scaled deployments.
+
+## API
+
+| Method | Endpoint | Purpose |
+|--------|----------|---------|
+| `POST` | `/api/v1/captures/upload` | Upload `.sub` files + GPS manifest |
+| `GET`  | `/api/v1/query/captures` | List stored captures |
+| `GET`  | `/api/v1/captures/{id}` | Capture detail |
+| `GET`  | `/api/v1/stats/summary` | Platform statistics |
+| `GET`  | `/api/v1/export?format=jsonl\|csv\|geojson` | Export the labeled dataset |
+| `GET`  | `/health` | Health check |
+
+## Supported formats
+
+Flipper Zero `.sub` — both decoded key files (`Protocol` / `Bit` / `Key`) and RAW pulse
+captures (`RAW_Data` timing arrays). `.fff` and PCAP ingest are on the roadmap.
+
+## Accuracy status
+
+Identification is at **Phase 0** — category routing plus a calibrated confidence model.
+The current gate benchmark (`scripts/benchmark_phase0.py`) reports **~67% top-3** category
+accuracy, but note this runs on **synthetic `.sub` files generated from the protocol database
+itself**, so treat it as an *upper bound*, not a real-world number. Real-world validation
+against field Flipper captures, and a hybrid heuristic + 1D-CNN ML ensemble, are the next
+milestones.
+
+## Tech stack
+
+Python · FastAPI · Leaflet.js · SQLAlchemy (PostgreSQL/PostGIS · SQLite) · Flipper `.sub` parsing
+
+## Roadmap
+
+- [x] `.sub` parser + GPS validation
+- [x] Category-routed device identification (Phase 0)
+- [x] Web UI (map / upload / search / stats) + dataset export
+- [ ] Real-world accuracy validation on field captures
+- [ ] ML ensemble (heuristic + 1D CNN on raw pulses + statistical features)
+- [ ] Community verification, accounts, leaderboard
+- [ ] Docker deployment; 2.4/5 GHz (PCAP + OUI) as a later phase