Split fuzzy search from the LLM path: /search/fuzzy and /geometry/{id}

[yellowcap]

Summary

• Adds GET /search/fuzzy — pure Jaro-Winkler place-name matching against divisions_area/natural_earth, no LLM. Returns the combined top-limit matches as a GeoJSON FeatureCollection (raw or simplified geometry via simplify). Pass ids_only=true to skip geometry fetching entirely and get back {"ids": [{"source", "id", "name", "country", "subtype", "admin_level", "bbox"}, ...]} — bbox ([minx, miny, maxx, maxy]) is computed via ST_XMin/YMin/XMax/YMax, far cheaper than full geometry.
• Adds GET /geometry/{id} — direct single-feature geometry lookup by ID (source inferred from the ne_ prefix if omitted), no fuzzy matching or LLM. Pairs with ids_only above for a cheap search-then-fetch workflow.
• Extracts geometry simplification/GeoJSON normalization out of sql.py into a standalone geometry.py with zero dependency on lm.py, so the fuzzy/ID-lookup paths never import DSPy or any LLM machinery — this is what actually enforces "no generative elements," not just a docstring claim.
• Loads the spatial extension once at startup via FastAPI lifespan onto a shared DuckDB connection (app.state.duckdb_con); request handlers use cheap cursor() calls instead of reconnecting per request. Removes ~90ms of redundant LOAD spatial overhead per call (previously paid twice on the simplify path — once in the endpoint, once hidden inside the geometry-normalization helper).
• gazet_demo.py: backend is now configurable via GAZET_BACKEND env var instead of hardcoded "gguf", so the demo can be pointed at either backend without editing source.

🐛 Bug fix: fuzzy matching was silently broken for most of divisions_area

Found while testing this PR manually — not introduced by this PR, but fixed here since it affects the new endpoints directly and shares code with the existing generative /search pipeline (both go through search_candidates):

• search_divisions_area matched against names.common.en, which is NULL for 824,502 of 1,071,266 rows (77%) of divisions_area. Since the query filtered on names.common.en IS NOT NULL, the large majority of admin divisions worldwide were silently invisible to every fuzzy search, in both /search/fuzzy and the existing /search LLM pipeline. Fixed by matching names.primary instead, consistent with how search_natural_earth already worked.
• Even for rows that did have a names.common.en value, an exact query like "Loja" scored worse against "Loja Province" (Jaro-Winkler 0.862) than against unrelated near-miss strings like "Lodja" (0.947), "Lojane" (0.933), "Loulja" (0.911) — so the real match ranked 4th+ instead of 1st. Fixed by adding a substring-match tier: candidates where the query is a literal substring of the name (contains(lower(name), lower(query))) are now ranked ahead of pure edit-distance near-misses, with similarity as the tiebreaker within each tier.
• Verified this also fixes real query results in the existing generative pipeline: /search?q=Loja&backend=dspy now correctly resolves to the actual Ecuador "Loja" region instead of a wrong candidate ("Lodja", a DR Congo county) that used to win purely on higher Jaro-Winkler score against the truncated names.common.en field.
• Also added country/subtype/admin_level to ids_only responses, since fixing the coverage bug surfaces multiple genuinely distinct real-world places sharing the same name (e.g. q=Loja now correctly returns Ecuador's "Loja" region, its nested "Loja" county, a "Loja" locality in Spain, and one in Latvia) — bare name alone can't disambiguate those.

Test plan

• ruff check clean on all touched files (pre-existing unrelated issues in sql.py left as-is)
• /search/fuzzy: verified geometry attached, sources filter, invalid-source 400, simplify=true/false size difference (14.6KB raw vs 4.5KB simplified), ids_only=true minimal response including country/subtype/admin_level/bbox
• /geometry/{id}: verified explicit source, auto-inference from ne_ prefix, 404 on missing ID, 400 on bad source
• /search and /search/stream (both gguf and dspy backends) still work end-to-end after the shared-connection refactor
• Confirmed per-request latency drop after the connection-reuse fix (~1200ms → ~980-1140ms for repeated /search/fuzzy calls)
• Confirmed the fuzzy-matching bug fix: q=Loja now returns all 5 real "Loja" divisions (Ecuador ×3 admin levels, Spain, Latvia) ranked at similarity 1.0, ahead of near-misses like "Aloja"; q=Georgia still returns exact matches first (now 5 instead of 4, since more of the dataset is reachable)