fix: closed segments no longer incorrectly remove turn permissions at non-adjacent intersections

[hamishcampbell]

Issue

Fixes #7332.

When --segment-speed-file closes a segment (speed 0), OSRM treats the whole edge-based node holding that segment as unusable. A start or end point on an open part of the same way, past the closure, either can't route at all or has to make a detour. For example, a house on segment A in the issue can't route north even though segment A itself is open.

This PR makes routing look at where the phantom sits relative to the closure, rather than only at whether the node has a closure somewhere on it. There are three commits, each one a separate case found & validated through manual testing:

1. fix(mld): start past a closure. The MLD customizer used to drop every turn leaving a node with a zero-speed segment. It now keeps those turns but marks them INVALID_EDGE_WEIGHT, so cell customization and through traffic still can't use them. In the MLD route, table and matching searches, a seed on such a node may leave it when no closed segment lies after the phantom (the seed weight shows this exactly). A source is now valid when its own segment is open. CH still drops these turns, so CH behaviour doesn't change for sources.

2. fix(engine): reach a target past a closure along its own edge. A target on a node with a closed segment before it used to be invalid outright, so a route from a source on the same node had to leave the node and U-turn back in from the other side. Now a phantom is a valid target, like a source, when its own segment is open. The searches refuse to reach it from the start of its node across the closure: a target seed can't be left past a closure before it, and the forward and reverse searches can only meet at such a seed when the source side entered the node past the closure. This covers CH and MLD route, table and alternative searches.

3. fix(engine): turn around at a waypoint heading towards a closure. With continue_straight a route that reached a waypoint heading towards a closure on the same node had nowhere to go, so the next leg failed. This is now treated like a dead end i.e the route may continue from the waypoint in the other direction, which continue_straight=false already allows.

Testing

• New cucumber scenarios in features/testbot/zero-speed-updates.feature:
  • Closing a segment should not trap a start past the closure on the same way (CH and MLD)
  • Turning around at a waypoint heading towards a closed segment
• features/testbot/snap_intersection.feature the b → e and d → a cases in "previous section of segment blocked" used to expect NoRoute. They now route under MLD (new @no_ch scenario). The a → e / e → a cases still give NoRoute.
• Manually checked against the NZ reproduction from the issue (Christchurch query, MLD, segment speed file) and it fixes the linked issue.

AI Disclosure

Claude was primarily used to find the fix this issue, alongside manual testing of routes and code review.

Tasklist

• self-review code for correctness and following the coding guidelines
• add tests (see testing)
• update relevant wiki pages (N/A I think)
• review
• adjust for comments

Requirements / Relations

N/A see the source issue #7332