feat(navigation): evaluate network detours for route stops

[Medformatik]

Summary

• Route-stop pins and cards now distinguish road-network detours from straight-line estimates, pending checks, unknown results and unreachable stops. A nearby stop across a river no longer appears as a short verified detour: the controlled screenshot fixture changes from about 1 minute to 10 minutes and 4 km by road.
• Compare the complete remaining itinerary with and without each shortlisted stop under the same routing mode, avoid preferences and provider. Preserve ordered user waypoints, including later visits to retraced roads, and prefer explicit entrances with a labelled coordinate fallback. Shared headless contracts provide the routing foundation for Combine fuel, charging and parking site evidence with along-route discovery #417.
• Bound evaluation to six candidates, two concurrent candidate requests and one baseline; cache up to 64 evaluations for 60 seconds at 500 m progress intervals. Cancel obsolete evaluations and selections, reject late route/session results, distinguish co-located stop identities and retain the provider chosen by the baseline. Update existing directions/navigation documentation and English/German copy. Existing routing administration applies; no new operator setting is required.

Related issues

Fixes #416

How was this tested?

• Independent directed graph fixtures cover motorway wrong-side access, a river crossing, an unreachable restricted stop and a reachable urban stop. Additional regressions cover entrance changes, repeated road visits, waypoint order, provider/option fidelity, request limits, expiry, cancellation, session replacement, refreshed selection coordinates and successful search completion.
• One independent whole-branch review identified four substantive edge cases; all were reproduced and fixed with red-to-green tests. Manual T3 preview verification identified and fixed search completion after a successful add. Final self-review reproduced and fixed a retained place ID using stale selection coordinates instead of its refreshed target; all 41 focused regressions passed.
• pnpm lint, pnpm check-types (30 workspaces plus discovery eval), pnpm check:policy, pnpm check-openapi, changeset status, the web production build and docs production build passed. The complete portable suite passed through the pre-push hook with four workers: 1,622 files / 17,840 tests passed, 30 files / 138 tests skipped by the standard opt-in gates. Concurrent full runs timed out unrelated checks; the publisher suite passed 18/18 in isolation and the final complete run passed with bounded workers, without changing unrelated code or test timeouts.
• Matched desktop and phone screenshots use the actual product control, hook and MapLibre results layer with controlled API/map fixtures. They demonstrate confidence rendering and interaction; they are not live-provider accuracy measurements. Screenshot assets are attached outside Git.

Checklist

• Conventional Commits title
• pnpm lint && pnpm check-types && pnpm test pass locally
• Changeset added for core and i18n
• Existing user and developer docs updated
• Code remains in the applicable license tiers
• No secrets, API keys or environment files committed
• Contributor License Agreement, if required by the CLA bot

Screenshots

View	Before	After
Desktop
Phone