Rewrite the backend in TypeScript with PostGIS location search

[savikthk]

Summary

Stage 1 of the plan: search by location that actually works, on a TypeScript backend instead of C++.

The old search found the place only through free-text Nominatim geocoding ("метро Тверская" landed on the Okhotny Ryad bus stop, so the 800 m circle moved almost a full radius), never used the user's own location, silently searched the whole city when geocoding failed and ranked venues inside the radius without looking at distance.

Now the algorithm does the work and GigaChat has a narrow role:

• Rules first: a vocabulary of categories, cuisines, moods, company, features and budgets plus an offline gazetteer from OpenStreetMap (240 metro stations, 667 landmarks, 132 district polygons), matched by Russian word stems.
• Search in PostGIS: 800 m around a station or landmark, the real district polygon, or 1 km around the user (Near me in the Mini App). When nothing fits, the radius widens to 1.5, 2.5 and 4 km before the category is relaxed, and notes in the response say what happened.
• Ranking by wishes, taste and distance, one branch per chain.
• GigaChat only interprets a query the rules did not understand and picks among the top 10 candidates the algorithm found (ids validated). After a network failure it is paused for two minutes and the app answers by the algorithm alone; Nominatim stays as a cached fallback with the same pause.

Changes

• backend/: Node 24, Fastify 5, zod, PostgreSQL 16 + PostGIS; migrations, rule parser and gazetteer, ranking, search and taste services, GigaChat and Nominatim adapters, places:fetch and db:load scripts. The C++ backend and load_to_db.py are removed.
• db/Dockerfile: PostgreSQL 16 with PostGIS that runs natively on Apple Silicon; docker-compose.yml builds it on the same data volume.
• frontend/: Near me button (Telegram LocationManager, browser geolocation outside Telegram), a line that explains where the search went, distances on cards, venue data rendered with textContent.
• CI: lint, typecheck and tests of the backend with a PostGIS service container instead of the C++ build.
• README: themed banner, badges, real query examples, a Mermaid diagram of the search.

Testing

• 84 unit and integration tests, the latter on a real PostGIS, npm run lint and npm run typecheck pass.
• Live on the full dataset: «бар у метро Тверская» gives bars 62–130 m from the station, «ресторан у Большого театра» restaurants from 109 m without the network, «кафе на Арбате» cafés inside the district, «боулинг у Тверской» widens to 2.5 km with a note, «кофейня рядом» with a location finds coffee 34 m away. GigaChat answered live once the network was back (ranked_by: "llm").

After merging, run docker compose up -d --build and cd backend && npm ci && npm run db:load once: the database image now includes PostGIS and the schema is migrated on start.