feat(data): POI build pipeline + chunked binary (PR 2 of 5)

[taddyb]

Summary

Second PR in the richer-geocoding rollout. Extracts ~80k categorized POIs from the same OSM PBF the walking graph already uses, pre-snaps each to a walk-graph node in the largest connected component (so future routing can't silently fail), and ships the result as a chunked binary.

No runtime code touched — this PR is data infrastructure. PR 3 will wire the POI binary into the search bar.

What's in the binary

• 79,863 POIs across the basemap (NYC + LI + south Westchester/Conn + NJ)
• 10 categories: food, transit, park, culture, attraction, shop, school, health, service, worship
• Each record carries a pre-snapped walk_node index (correctness win — fixes the class of bug where a POI lands inside an OSM-island park path and becomes unrouteable)
• File size: 2.5 MB → single chunk tiles/pois.part-aa (well under the 50 MB / GitHub 100 MB cap)

New files

Path	Purpose
pipelines/walk_graph_reader.py	Decode walk_graph.bin (NWLK format) for offline tools; LCC + local snap
pipelines/poi_emit.py	POI1 binary writer (24B header, 20B records, names blob)
pipelines/pois.py	pyosmium extractor: classify tags → category, pre-snap, clip to bbox
scripts/build-pois.sh	Bash wrapper; runs the pipeline + chunks to tiles/pois.part-*
scripts/assemble-pois.sh	Idempotent reverse: cat chunks → tiles/pois.bin
tiles/pois.part-aa	Committed POI chunk
pipelines/tests/test_*.py	8 new pytest cases (binary round-trip, classify, extract, LCC snap)

CI

• pages.yml now assembles pois.bin from chunks alongside the walk-graph
• pr-checks.yml gains a pipeline-tests job that runs uv run pytest against the assembled walk-graph

Test plan

• `uv run --directory pipelines pytest -v` → 15/15 pass locally
• `bash scripts/build-pois.sh` produces a sane binary (79,863 POIs / 2.5 MB)
• `bash scripts/assemble-pois.sh` round-trips chunks back to identical `pois.bin`
• CI's new `pipeline-tests` job runs pytest with the assembled walk graph
• CI's existing deploy job assembles `pois.bin` without error

Subsequent PRs

• PR 3: local POI lane in the search dropdown (consumes `pois.bin` at runtime)
• PR 4: WASM `reachable_nodes` method
• PR 5: search-driven category discovery + map pins

🤖 Generated with Claude Code