feat: offer Nominatim entrance nodes as routing targets

[DennisOSRM]

A geocoded place drops its waypoint on the centroid. For a large site that is nowhere a traveller can reach — BER's centroid sits on the airfield, 1.5 km from the terminal doors, and the route ends on a slip road — Nominatim's entrance nodes are now offered as an alternative.

Clicking a door routes to it while the pin stays on the place that was searched for, joined by a dashed line. Every door stays on the map, the chosen one marked; clicking it again routes back to the place.

	before	after
BER, driving	23.1 km, 34 min, ends on the A 113 slip road	24.5 km, 29 min, at the terminal door
Pergamonmuseum, walking	1.6 km, 22 min, behind the museum	1.5 km, 20 min, at the entrance

Nothing is chosen for the user

entrance=main records how a building is laid out, not what a router can reach. The Pergamonmuseum tags a main door that the walking network stops 90 m short of, while its side door is on the path — so acting on the tag would be confidently wrong about as often as right. Every door is offered; the waypoint stays put until one is clicked.

Which doors are offered

Direction, from the wiki's one-way values:

value	wiki	usable at
exit	"a one-way out of a building"	origin only
entrance	"an entrance only, a one-way in"	destination only
main, yes, secondary, shop, home	enter or leave	anywhere
service, emergency, staircase, garage, no	—	never

A via point is both arrived at and left from, so it takes only the two-way doors. Hardenbergstraße 36 has six mapped doors and offers 5 / 5 / 4 as origin / destination / via.

Travel mode, from the door's own OSM access tags, following the access hierarchy (motor_vehicle → vehicle → access; foot → access). Only no and private forbid — permissive, customers and designated are usable, and an untagged door is open to everyone. Read live, so switching profile re-applies the rules to what is already on screen.

Labels

Doors take their OSM name where they have one (Alexa Berlin: Haupteingang Alexanderplatz, Eingang Ravelinplatz…), shown permanently rather than on hover. Where labels collide the whole colliding run collapses into one listing every door in it; zooming in separates them again. Overlap is transitive — showing A and C while hiding B would be arbitrary.

Labels sit in a pane below the marker pane. A zIndexOffset cannot do this: Leaflet derives a marker's z-index from its latitude, so a label on a northerly door outranks a dot on a southerly one however the offsets are set.

Also here

• Site outline beneath the doors, fetched only when the picker opens. Thinned with Visvalingam–Whyatt, stopping as soon as the smallest surviving triangle would be visible on screen, rather than with Nominatim's polygon_threshold — that tolerance is absolute degrees, so a value that usefully thins a 5 km airport flattens a 170 m building into five stray corners. Measured: the museum keeps 31 of 32 points, BER sheds 33%, neither bounding box moves by a pixel.
• The gap between where a route can reach and its waypoint is dashed throughout. LRM's default draws it as a solid black casing and solid white core with thin dashes on top, which against this app's heavy blue route reads as more route.
• continue_straight=true, so OSRM leaves a via by a different way rather than doubling back — 1575 m against 1524 m on a via at Hardenbergstraße 36. A no-op without via points, and it still routes where the u-turn is unavoidable.
• Cache key → v2. Existing entries predate the entrance data, the OSM identifiers and the access tags.

Tests

465 tests, 32 suites. New modules are at 100% line coverage (entrance_picker, entrance_waypoints, simplify, lrm_options); the remaining branch misses are defensive fallbacks. The entrance-specific geocoder paths — both request paths, forward and reverse, and the persisted-cache round trip — are covered too.

Verified against the live Nominatim and OSRM services throughout, not only against fakes.

Known gap

A waypoint that becomes a via — by appending a waypoint after it — keeps a door the via rules would not have offered. The role changes; the existing selection is not re-validated. Re-typing the address recovers it. Fixing it properly means retaining per-waypoint place state across splices, which the picker deliberately does not keep today, so I have left it rather than bolt on a half-measure.

🤖 Claude Code, Claude Opus 5