Let features own their navigation graphs

[AhmadKharfan]

Why

Every route in the app was declared in one Routes object in :app, and NavGraphs.kt built every destination itself, importing screen composables from all seven feature modules. A feature could not describe its own navigation; adding a screen meant editing :app.

What changed

Four features now own their routes and register their own destinations:

• :feature:onboarding — OnboardingRoutes + onboardingGraph
• :feature:settings — SettingsRoutes + settingsGraph
• :feature:projects — ProjectsRoutes + projectsGraph
• :feature:terminal — TerminalRoutes + terminalGraph

:app keeps the single NavHost and calls those graphs. Cross-feature navigation stays as callbacks supplied by :app, so no feature learns another feature's routes: onboarding takes an onFinished, projects takes onOpenProject and onOpenPreferences.

encodeRouteArg moved to :core:common so any module building a route can escape its arguments, rather than the helper being private to :app.

ProjectsRoutes also owns the picked_folder saved-state key. Its producer (the folder picker) and both consumers (hub, create-project) are all inside :feature:projects, so the protocol belongs there. The git_conflict_path key deliberately stays in :app — its producer and consumer are in different features, and moving it would make one import a constant from the other, recreating the coupling at the navigation layer.

Behavior

No behavior change, and this was checked rather than assumed: all 15 relocated route strings are byte-identical to their previous values, verified by diffing the moved constants against Routes.kt at the previous commit. Destination registration order, argument extraction and composable wiring are unchanged; only the file that declares them moved.

Scope

Two deliberate exclusions.

Type-safe routes are not part of this change. The plan listed converting from string routes to type-safe routes as a preparatory step. It changes route matching and argument parsing, and the only meaningful verification for that is on a device. Keeping the strings identical is what makes this change provably behaviour-preserving, so the conversion is better done separately.

The editor and git destinations stay in :app for now. They are entangled: the editor's navigation callbacks target git destinations, and git's conflict screen hands a path back to the editor. Splitting them is part of the change that eliminates the editor-to-git dependency, where the contract between them is being designed anyway.

Tests

• 714 tests passing, 0 failures.
• RoutesTest continues to cover the route builders that remain in :app.

Verification

From a clean worktree at the branch tip:

• ./gradlew test detekt verifyModuleBoundaries :app:assembleDebug — BUILD SUCCESSFUL
• 714 tests, 0 failures
• All 15 moved route strings confirmed identical to the previous commit

What is not verified here: back-stack behaviour, saved-state result delivery across process death, and predictive back are only observable on a device, and this change was not exercised on one. The byte-identical route strings and unchanged wiring are the argument that behaviour is preserved; a device pass over hub → create project → folder picker → back, and onboarding → hub, is still worth doing before release.

Note on included work

This branch also carries the route-argument encoder and its tests, which existed as uncommitted work in the tree from a separate task. Phase 4 needed a shared encoder for features to build routes, so it was moved to :core:common rather than left stranded in :app. The implementation and its tests are unchanged apart from the package.