feat: playground & restroom amenity POI types (#418)

[fatherlinux]

Summary

Adds Playground and Restroom as map POI types so parents can find parks with good amenities, seeds real locations from OpenStreetMap, and excludes these amenity types from news/events collection.

• Types (migration 066): playground + restroom icon rows. Name-classified with no activity fallbacks, so a full park that merely lists "Playground" among its activities keeps its own icon. Lowest sort_order so the explicit restroom/playground keyword beats park-name keywords (e.g. mill→historic).
• icons.default_hidden: amenities appear in the legend but start toggled off to avoid clutter; users opt in.
• pois.osm_id: provenance + idempotency key (unique partial index).
• OSM data: backend/data/osm/amenities.json — 240 features inside boundary_type='park' polygons (201 restrooms, 39 playgrounds). Idempotent importer import-osm-amenities.js upserts point POIs by osm_id, named from the OSM tag or "<Park> Restroom/Playground".
• Collection skip: news_collection_excluded_types (default ["playground","restroom"]); backend/utils/poiClassify.js + the two selection helpers drop excluded types (by name classification or sole-activity match). Verified 0 amenities leak; real parks still collect.
• Attribution: OSM/ODbL credit added to the satellite base layer (street layer already credits OSM).

No pois-schema change beyond the additive osm_id column; 066 is idempotent.

Closes #418

Deploy steps (beyond image pull + restart)

1. 066_add_amenity_poi_types.sql runs automatically on container init.
2. Run the importer once: node /app/migrations/import-osm-amenities.js.

Test plan

• ./run.sh build (frontend compiled via reload rebuild)
• Legend shows Playground/Restroom, default off; toggling reveals pins inside parks
• Amenity POIs have no News/Events tabs; 0 leak into collection; real parks still collect
• "Valley View Woods Park" keeps its normal icon (regression fixed)
• Importer idempotent (re-run = 0 inserts); 8/8 poiClassify unit tests; gourmand clean
• Human verification in browser

🤖 Generated with Claude Code

[fatherlinux]

Gatehouse review disposition

Fixed (doc, 1 commit):

• MED: documented that classifyPoiType's iconConfig must be ordered by sort_order ascending (the implicit priority contract).
• LOW: commented the xmax = 0 insert/update sentinel in the importer.

Justified — false positives:

• 2× CRITICAL ("primary_activities is a TEXT[] array → .trim()/.toLowerCase() TypeError"): pois.primary_activities is text, not text[] (confirmed via information_schema: data_type=text; sample value "Hiking, Picnicking, …"). The pg driver returns a plain comma-separated string, so the string methods are correct. The live collection run executed without error and classified correctly (0 amenities leaked, real parks still selected) — which it could not do if the value were an array.

Justified — no change:

• MED (N+1 upsert in importer): one-time import of 240 rows, run manually at deploy; a bulk statement adds complexity for no meaningful gain at this size.
• MED (sort_order overloaded for display + classification): pre-existing semantics — the frontend classifier already iterates icons by sort_order; this PR documents it in the migration but doesn't introduce the dual use. Legend display sorts by label separately.
• MED (iconConfig loaded per filterExcludedTypePois call): called once per collection job (not per POI), so it's one extra cheap query per batch run.

[gemini-code-assist]

Code Review

This pull request introduces "Playground" and "Restroom" as first-class map POI types, seeds them using a committed OpenStreetMap snapshot, and excludes these amenity types from news/events collection to save API budget. Feedback was provided to optimize the backend POI classifier by caching compiled regular expressions in matchesWholeWord to prevent performance degradation during batch processing.

[gemini-code-assist]

Compiling a new RegExp object on every call to matchesWholeWord inside nested loops for every POI can lead to significant performance degradation during batch processing of hundreds or thousands of POIs. Caching the compiled regular expressions by keyword avoids redundant compilation overhead.

Suggested change

	function matchesWholeWord(text, keyword) {
	const escaped = keyword.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
	return new RegExp(`\\b${escaped}\\b`, 'i').test(text);
	}
	const regexCache = new Map();
	function matchesWholeWord(text, keyword) {
	let regex = regexCache.get(keyword);
	if (!regex) {
	const escaped = keyword.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\$&');
	regex = new RegExp('\\\\b' + escaped + '\\\\b', 'i');
	regexCache.set(keyword, regex);
	}
	return regex.test(text);
	}