refactor: organize app.state into typed namespaces

[antosubash]

Summary

Consolidates 15+ loose app.state.* attributes into one typed slot per owner: framework gets app.state.sm: Services (a frozen dataclass), each module with state gets app.state.<module>: <Module>Services. Drops three duplicate locale fields and runtime drift-detection. Storage stays on app.state — only its shape changes.

Full design and rationale: docs/superpowers/specs/2026-04-17-app-state-organization-design.md.
Step-by-step execution plan: docs/superpowers/plans/2026-04-17-app-state-organization.md.

What changed

New API

• simple_module_core.Services — frozen + slotted dataclass, 10 fields (settings, db, event_bus, menu_registry, permissions, feature_flags, health_registry, i18n_registry, inertia_config, modules).
• users.services.UsersServices — module-scoped container for settings + mailer + rate limiters + roles cache.
• Scaffolding now emits a services.py per new module via make new-module.

Kept loose on app.state (intentional, documented)

• app.state.inertia_dependency — fastapi-inertia's request-scoped Depends factory.
• app.state.migration — dev-only boot-time check result set in lifespan.

Deleted

• app.state.settings_default_locale, settings_supported_locales, settings_cookie_name — pure duplicates of settings.i18n_*.
• The state_before/state_after drift-detection block in create_app that used to back SM012.
• app.state.inertia_config loose write — setup_inertia now returns the config; create_app passes it directly into Services(...).

Diagnostic update

• SM012 re-aimed: warns when register_settings is overridden but app.state.<module_lower> wasn't populated.

Rollout shape

Shipped in 19 commits, each independently revertable, keeping main green at every boundary:

1. Introduce Services + populate app.state.sm alongside existing writes.
2. Migrate framework consumers (middleware, _error_handlers, health, inertia_deps, i18n_deps, permissions, lifespan block).
3. Migrate module consumers (users module settings, mailer, rate limiters, roles cache, bootstrap).
4. Delete the loose keys + duplicates + drift detection.
5. Update scaffolding templates + docs/framework-conventions.md.

Docs-to-CLI check: make doctor still clean; no new framework diagnostics.

Test plan

• make ci-python-lint — ruff format + check, 205 files clean
• make ci-python-typecheck — ty, all checks passed
• make ci-js-lint — biome, 117 files clean
• make ci-js-typecheck — tsc for client app + 5 tsconfig projects
• make ci-check-file-size — no files exceed 300 lines
• make test-py — 562 passed, 4 deselected
• make test-js — 8 passed
• make doctor — exit 0, zero diagnostics
• make build — vite build green
• make test-e2e — not run locally (needs Docker + dev server + uv run playwright install chromium); CI will exercise

Notes for reviewers

• The migration touches ~40 call-sites but the mechanical shape is uniform: request.app.state.<x> → request.app.state.sm.<x> (framework) or .users.<x> (module). Most commits are small and read straightforwardly.
• Non-frozen UsersServices is deliberate — on_startup sets mailer/rate_limiter/roles_cache after register_settings. Convention (documented in framework-conventions.md) says "set once during boot, read-only after" without enforcement.
• The scaffolding templates include the new pattern, so any make new-module name=... today adopts it out of the box.