Skip to main content

umbral_core/
app.rs

1use axum::Router;
2use std::collections::HashMap;
3use std::net::SocketAddr;
4
5use crate::db::{self, DbPool};
6use crate::migrate::ModelMeta;
7use crate::orm::Model;
8use crate::plugin::Plugin;
9use crate::settings::Settings;
10
11/// A per-request resolver that builds the request-scoped
12/// [`crate::db::RouteContext`] from the incoming request. Installed via
13/// [`AppBuilder::route_context`] and driven by [`route_context_scope_layer`].
14type RouteContextResolver =
15    std::sync::Arc<dyn Fn(&crate::web::Request) -> crate::db::RouteContext + Send + Sync>;
16
17/// A built and ready-to-serve umbral application.
18///
19/// Created via `App::builder().build()`. Owns the merged router that
20/// carries every registered plugin's routes plus the user-binary
21/// routes passed to `AppBuilder::routes()`.
22pub struct App {
23    router: Router,
24    plugins: Vec<Box<dyn Plugin>>,
25    /// Management commands the *project* registered directly, via
26    /// [`AppBuilder::command`] — the ones that belong to the binary
27    /// rather than to any plugin. Handed to
28    /// [`crate::cli::dispatch_with_app_commands`] alongside the plugins'
29    /// own contributions.
30    commands: Vec<Box<dyn crate::cli::PluginCommand>>,
31    /// gaps3 #23: when true, `umbral_cli::dispatch` applies pending migrations
32    /// before starting the server (the `serve` command only) — so a fresh DB
33    /// "just works" WITHOUT running migrate during `makemigrations`/`migrate`
34    /// or any other subcommand. Opt in via [`AppBuilder::auto_migrate_on_serve`].
35    auto_migrate_on_serve: bool,
36    /// Kikosi #5 — how long [`App::serve`] keeps serving after a shutdown signal
37    /// before it stops accepting, so a load balancer observes `/readyz` flip to
38    /// 503 and drains this instance. `Duration::ZERO` (the default) skips the
39    /// drain — the historical behaviour. Set via [`AppBuilder::shutdown_drain`].
40    drain_delay: std::time::Duration,
41    /// Set the first time [`App::ready`] fires the `on_ready` hooks, so the
42    /// second call is a no-op. `serve()` and `into_router()` both call it, and
43    /// `umbral_cli::dispatch` may have called it already.
44    ready_fired: std::sync::atomic::AtomicBool,
45}
46
47impl App {
48    /// Whether the app opted into auto-migrate on `serve` (gaps3 #23). Read by
49    /// `umbral_cli`'s serve path; see [`AppBuilder::auto_migrate_on_serve`].
50    pub fn auto_migrate_on_serve_enabled(&self) -> bool {
51        self.auto_migrate_on_serve
52    }
53
54    /// Fire every plugin's [`Plugin::on_ready`] hook, in topological order.
55    /// Idempotent: the second and later calls do nothing.
56    ///
57    /// # Why this is not part of `build()`
58    ///
59    /// `on_ready` means *the application is up*. Plugins use it to seed content,
60    /// backfill rows, install RLS policies, and (in `umbral-permissions`) create
61    /// the standard permission rows for every registered model. All of that
62    /// needs a migrated schema.
63    ///
64    /// [`AppBuilder::build`] still calls this for you, so a test or an embedder
65    /// that holds an `App` sees no change. What changed is `umbral_cli::dispatch`:
66    /// it takes the *builder*, calls [`AppBuilder::build_deferred`], resolves
67    /// argv, and only then calls `ready()` — skipping it entirely for the schema
68    /// commands (`migrate`, `makemigrations`, `inspectdb`, …).
69    ///
70    /// Before that, the generated `main.rs` was
71    /// `let app = App::builder()…build()?; umbral_cli::dispatch(app)`, so the
72    /// hooks ran before `dispatch` had even parsed argv — including when argv
73    /// said `migrate`. Against a fresh database that produced a wall of
74    /// `relation "…" does not exist` before the migration engine had created a
75    /// single table (gaps3 #41, seen on the first umbralrs.dev deploy). Nothing
76    /// crashed only because those seeds log-and-swallow; a plugin that propagated
77    /// the error made `migrate` unrunnable, and one that wrote rows silently
78    /// skipped the write.
79    ///
80    /// [`App::serve`] calls this too, so a hand-rolled `main` that builds with
81    /// `build_deferred()` and serves directly still gets its hooks.
82    /// Whether [`App::ready`] has already run the `on_ready` hooks.
83    ///
84    /// `umbral_cli::dispatch` reads this to warn when a binary still builds with
85    /// `App::build()` — the hooks fired before argv was parsed, so a schema
86    /// command has already run every plugin's seed (gaps3 #41).
87    pub fn ready_already_fired(&self) -> bool {
88        self.ready_fired.load(std::sync::atomic::Ordering::SeqCst)
89    }
90
91    pub fn ready(&self) -> Result<(), BuildError> {
92        use std::sync::atomic::Ordering;
93        if self.ready_fired.swap(true, Ordering::SeqCst) {
94            return Ok(());
95        }
96        let ctx = crate::plugin::AppContext {
97            pool: crate::db::pool_dispatched().clone(),
98            settings: crate::settings::get().clone(),
99        };
100        for plugin in &self.plugins {
101            plugin
102                .on_ready(&ctx)
103                .map_err(|source| BuildError::PluginOnReady {
104                    plugin: plugin.name(),
105                    source,
106                })?;
107        }
108        Ok(())
109    }
110
111    /// Create a new [`AppBuilder`].
112    pub fn builder() -> AppBuilder {
113        // Load `.env` into the *process* environment so plain
114        // `std::env::var(...)` code sees it — most importantly a plugin's
115        // `from_env()` credential loader (e.g. the OAuth providers reading
116        // `UMBRAL_OAUTH_*`). This runs before the `.plugin(...)` arguments
117        // are evaluated, so those loaders find the values.
118        //
119        // We read `.env` the *same* CWD-relative way figment's settings
120        // loader does (`from_filename_iter(".env")`) rather than
121        // `dotenvy::dotenv()`, whose parent-directory search resolves the
122        // file differently and missed it in practice. Each key is set only
123        // when it isn't already present, so real environment vars keep
124        // precedence. No-op when there's no `.env`.
125        if let Ok(iter) = dotenvy::from_filename_iter(".env") {
126            for (key, value) in iter.flatten() {
127                if std::env::var_os(&key).is_none() {
128                    // SAFETY: runs at startup (App::builder), before the
129                    // server spawns request handlers that read the
130                    // environment — the same operation `dotenvy::dotenv()`
131                    // performs internally.
132                    unsafe { std::env::set_var(&key, &value) };
133                }
134            }
135        }
136        AppBuilder::default()
137    }
138
139    /// Bind the axum listener and serve requests.
140    ///
141    /// Fires [`App::ready`] first (idempotent, so `umbral_cli::dispatch` having
142    /// already called it is fine): a server that is about to accept requests is
143    /// by definition ready, and a plugin's `on_ready` may install the ambient
144    /// state its handlers read. A hook that fails surfaces as an
145    /// [`std::io::ErrorKind::Other`] carrying the `BuildError`'s message —
146    /// `serve` has always returned `io::Error`, and a plugin that can't start is
147    /// as fatal as a port that won't bind.
148    ///
149    /// This call blocks until the server stops. At M0 there is no graceful
150    /// shutdown hook; that lands with the signal-handling work in a later
151    /// milestone.
152    pub async fn serve(self, addr: impl Into<SocketAddr>) -> Result<(), std::io::Error> {
153        self.ready().map_err(std::io::Error::other)?;
154
155        let listener = tokio::net::TcpListener::bind(addr.into()).await?;
156
157        tracing::info!("umbral serving on {}", listener.local_addr()?);
158
159        // Serve via `into_make_service()` rather than passing the router
160        // directly. `axum::serve(listener, router)` drives the `Router` as
161        // its own connection-maker, whose per-connection `call` runs
162        // `self.clone().with_state(())` — and `with_state` finalizes EVERY
163        // route eagerly, an O(route-count) cost paid once per new TCP
164        // connection. With keep-alive that's amortized over all requests on
165        // the connection; WITHOUT keep-alive (one connection per request) it
166        // is paid on every request, capping throughput at ~1/with_state-cost
167        // regardless of the handler. For an app with hundreds of routes (a
168        // full admin + REST surface) that throttled no-keep-alive throughput
169        // by ~4x or worse. `IntoMakeService` instead hands each connection a
170        // cheap `Router::clone()` (an `Arc` bump) and lets routing finalize
171        // lazily per request — measurably faster on fresh connections and no
172        // slower with keep-alive. No `ConnectInfo` regression: the direct
173        // path didn't provide it either (that needs
174        // `into_make_service_with_connect_info`).
175        // audit_2 core-app-config #13: graceful shutdown. Without it, a deploy
176        // (SIGTERM) drops every in-flight request and never drains the pools —
177        // Postgres logs abrupt terminations, SQLite skips its WAL checkpoint.
178        // `with_graceful_shutdown` stops accepting new connections on the
179        // signal and waits for in-flight requests to finish; then we close the
180        // pools so connections shut down cleanly.
181        // Kikosi #5: when a drain delay is configured, the shutdown future flips
182        // readiness to draining and holds for the delay BEFORE resolving, so the
183        // server keeps accepting during the window the load balancer needs to
184        // notice `/readyz` = 503 and stop routing here. With ZERO delay this is
185        // the plain signal wait — the historical behaviour.
186        let drain_delay = self.drain_delay;
187        axum::serve(listener, self.router.into_make_service())
188            .with_graceful_shutdown(drain_after(shutdown_signal(), drain_delay))
189            .await?;
190        tracing::info!("umbral: server stopped accepting; draining DB pools");
191        crate::db::close().await;
192        Ok(())
193    }
194
195    /// Consume the [`App`] and return its merged axum router.
196    ///
197    /// Useful when the caller wants to drive the router themselves: an
198    /// integration test that sends synthetic requests via
199    /// `tower::ServiceExt::oneshot`, an embedding scenario that nests
200    /// umbral under another axum tree, or any other path that doesn't
201    /// want `serve()`'s opinionated listener.
202    pub fn into_router(self) -> Router {
203        self.router
204    }
205
206    /// Borrow the registered plugins in topological dependency order.
207    ///
208    /// Used by [`crate::cli::dispatch`] to walk every plugin's
209    /// `commands()` contribution at CLI dispatch time. Borrowed (not
210    /// moved) so the App stays usable after a dispatch call returns.
211    pub fn plugins(&self) -> &[Box<dyn Plugin>] {
212        &self.plugins
213    }
214
215    /// Borrow the project's own commands — the ones registered directly on
216    /// the builder via [`AppBuilder::command`] rather than contributed by a
217    /// plugin.
218    ///
219    /// Mirrors [`App::plugins`]: borrowed, not moved, so the App stays
220    /// usable after [`crate::cli::dispatch_with_app_commands`] returns
221    /// `Unmatched` and the caller falls through to its built-ins.
222    pub fn commands(&self) -> &[Box<dyn crate::cli::PluginCommand>] {
223        &self.commands
224    }
225}
226
227/// The fluent entry point for constructing an [`App`].
228///
229/// Collects settings, database pools, and routes, then locks everything
230/// into place at [`build`](AppBuilder::build).
231pub struct AppBuilder {
232    settings: Option<Settings>,
233    databases: HashMap<String, DbPool>,
234    router: Option<Router>,
235    /// Companion path list for `router` — surfaces the user's hand-
236    /// registered routes in the dev-mode 404 page. The builder can't
237    /// peek inside an axum `Router`, so the caller declares its paths
238    /// here. Empty by default; production deployments don't need to
239    /// fill it.
240    route_paths: Vec<crate::routes::RouteSpec>,
241    models: Vec<ModelMeta>,
242    /// gaps3 #46 — collect link-registered models at build time.
243    auto_models: bool,
244    plugins: Vec<Box<dyn Plugin>>,
245    /// Project-owned management commands, added via [`AppBuilder::command`].
246    /// Kept out of the plugin list on purpose: a command the binary owns
247    /// isn't a reusable unit, and wrapping it in a dummy plugin to reach
248    /// argv would be a workaround for a missing contract, not a design.
249    commands: Vec<Box<dyn crate::cli::PluginCommand>>,
250    templates_dir: Option<std::path::PathBuf>,
251    slash_redirect: crate::slash::SlashRedirect,
252    not_found_template: Option<String>,
253    server_error_template: Option<String>,
254    /// Custom template per status code for general error pages (429, 403, …),
255    /// styled like the 404/500 pages. See [`Self::error_template`].
256    error_templates: HashMap<axum::http::StatusCode, String>,
257    /// Optional hook called before the 500 template is rendered.
258    server_error_hook: Option<crate::errors::ServerErrorHook>,
259    /// When `true` (the default), the embedded default 404/500 templates
260    /// are used as fallbacks when the user hasn't supplied their own.
261    default_error_pages: bool,
262    /// gaps3 #23: apply pending migrations on `serve` (opt-in).
263    auto_migrate_on_serve: bool,
264    /// Kikosi #5: shutdown drain delay. `Duration::ZERO` = no drain.
265    drain_delay: std::time::Duration,
266    /// Path-scoped cross-origin policies (prefix → config), applied via
267    /// [`AppBuilder::cors_for`]. Each is layered only onto requests whose
268    /// path starts with the prefix (e.g. `"/api"`).
269    cors_scoped: Vec<(String, crate::cors::CorsConfig)>,
270    /// Optional cross-origin policy. `None` means no `CorsLayer`
271    /// is installed at all and browsers apply the same-origin
272    /// default. Configure via [`AppBuilder::cors`].
273    cors: Option<crate::cors::CorsConfig>,
274    /// When `Some(true)`, every ORM write terminal that supports
275    /// `.atomic()` / `.non_atomic()` runs inside a transaction by
276    /// default. Per-call `.non_atomic()` overrides. `None` keeps the
277    /// pre-flag behaviour (no auto-wrapping). See
278    /// [`AppBuilder::atomic_transactions`].
279    atomic_transactions: Option<bool>,
280    /// When `true`, a `tower-http` gzip/brotli compression layer wraps the
281    /// router. Off by default — a reverse proxy usually owns compression,
282    /// and double-compressing behind one is wasteful. Enable via
283    /// [`AppBuilder::compression`].
284    compress: bool,
285    /// Framework-wide request-body size cap (bytes). `build()` installs a
286    /// `tower-http` `RequestBodyLimitLayer` so any body over the cap is
287    /// rejected with `413` before a handler buffers it (audit_2 core-web H11).
288    /// Defaults to 32 MiB; `None` disables the global limit. Set via
289    /// [`AppBuilder::max_request_body`].
290    max_request_body_bytes: Option<usize>,
291    /// Per-request timeout. `build()` installs a `tower-http` `TimeoutLayer`
292    /// so a hung/slowloris request is aborted with `408` instead of pinning a
293    /// task forever (audit_2 core-web H11/#3). Defaults to 30s; `None`
294    /// disables. Set via [`AppBuilder::request_timeout`].
295    request_timeout: Option<std::time::Duration>,
296    /// Ship minimal hardening response headers from core (audit_2 H10):
297    /// `X-Content-Type-Options: nosniff`, `X-Frame-Options: DENY`,
298    /// `Referrer-Policy: strict-origin-when-cross-origin` — set ONLY if not
299    /// already present, so `SecurityPlugin` (which owns the configurable values
300    /// + CSRF + HSTS) wins when mounted. Default `true`; opt out via
301    /// [`AppBuilder::default_security_headers`].
302    default_security_headers: bool,
303    /// App-level framework middleware (feature #68), prepended to the
304    /// plugins' contributions in the final stack. Added via
305    /// [`AppBuilder::middleware`].
306    middleware: Vec<std::sync::Arc<dyn crate::middleware::Middleware>>,
307    /// Optional custom [`crate::db::DatabaseRouter`]. `None` uses
308    /// `DefaultRouter` (today's static per-model routing). Installed
309    /// during `build()` via [`crate::db::router::install_router`].
310    db_router: Option<std::sync::Arc<dyn crate::db::DatabaseRouter>>,
311    /// Optional per-request resolver that builds the request-scoped
312    /// [`crate::db::RouteContext`]. When set, `build()` installs a layer that
313    /// runs the resolver on each request and scopes the ENTIRE downstream
314    /// future (handler plus every `.await`, including ORM calls) inside
315    /// [`crate::db::route_context::scope`], so the ambient
316    /// `umbral::db::route_context()` accessor — and thus the `DatabaseRouter`
317    /// — sees the context this resolver set. Added via
318    /// [`AppBuilder::route_context`].
319    route_context_resolver: Option<RouteContextResolver>,
320    /// When `true`, `build()` FAILS (not just warns) if any app-level mutating
321    /// route (POST/PUT/PATCH/DELETE) registered via `.routes(...)` carries no
322    /// recorded permission (gaps3 #28 P1 — enforces the audit_2 H19 audit).
323    /// Opt-in "gated by construction": a forgotten authorization gate becomes a
324    /// boot error instead of a silently-open endpoint. Default `false` (warn
325    /// only). Set via [`AppBuilder::deny_ungated_mutations`].
326    deny_ungated_mutations: bool,
327}
328
329impl Default for AppBuilder {
330    fn default() -> Self {
331        Self {
332            settings: None,
333            databases: HashMap::new(),
334            router: None,
335            route_paths: Vec::new(),
336            models: Vec::new(),
337            auto_models: false,
338            plugins: Vec::new(),
339            commands: Vec::new(),
340            templates_dir: None,
341            slash_redirect: crate::slash::SlashRedirect::default(),
342            not_found_template: None,
343            server_error_template: None,
344            error_templates: HashMap::new(),
345            server_error_hook: None,
346            default_error_pages: true,
347            auto_migrate_on_serve: false,
348            drain_delay: std::time::Duration::ZERO,
349            cors: None,
350            cors_scoped: Vec::new(),
351            atomic_transactions: None,
352            deny_ungated_mutations: false,
353            compress: false,
354            // Safe-by-default request hardening (audit_2 core-web H11): a 32
355            // MiB body ceiling and a 30s timeout, both opt-out-able.
356            max_request_body_bytes: Some(32 * 1024 * 1024),
357            request_timeout: Some(std::time::Duration::from_secs(30)),
358            default_security_headers: true,
359            middleware: Vec::new(),
360            db_router: None,
361            route_context_resolver: None,
362        }
363    }
364}
365
366impl AppBuilder {
367    /// Set the application settings.
368    pub fn settings(mut self, settings: Settings) -> Self {
369        self.settings = Some(settings);
370        self
371    }
372
373    /// Register a database pool under the given alias.
374    ///
375    /// The `"default"` pool is the one returned by `umbral::db::pool()`
376    /// and is required: `build()` fails with `BuildError::
377    /// DefaultPoolMissing` if it isn't registered. The caller opens
378    /// the pool via `umbral::db::connect(&url).await` and passes it
379    /// here.
380    ///
381    /// Accepts anything that converts into a [`DbPool`]: a typed
382    /// [`sqlx::SqlitePool`], a typed [`sqlx::PgPool`], or an already-
383    /// built `DbPool`. The [`From`] impls on `DbPool` make plain
384    /// SqlitePool callers (every test, every plugin example) work
385    /// unchanged.
386    pub fn database(mut self, alias: &str, pool: impl Into<DbPool>) -> Self {
387        self.databases.insert(alias.to_owned(), pool.into());
388        self
389    }
390
391    /// Install a custom [`crate::db::DatabaseRouter`]. Omit to use
392    /// `DefaultRouter` (today's static per-model routing).
393    pub fn router<R: crate::db::DatabaseRouter + 'static>(mut self, router: R) -> Self {
394        self.db_router = Some(std::sync::Arc::new(router));
395        self
396    }
397
398    /// Install a per-request [`crate::db::RouteContext`] resolver.
399    ///
400    /// The resolver runs once per request, builds a `RouteContext` (typically
401    /// reading a tenant header or subdomain), and `build()` wraps the entire
402    /// downstream future in [`crate::db::route_context::scope`]. Because the
403    /// scope spans the whole handler — including every `.await` and every ORM
404    /// call — the ambient `umbral::db::route_context()` accessor inside the
405    /// handler, and the active [`crate::db::DatabaseRouter`], see exactly the
406    /// context this resolver returned. A request the resolver maps to a
407    /// default `RouteContext` runs with no tenant (no silent inheritance from
408    /// a prior request).
409    ///
410    /// ```ignore
411    /// use umbral::prelude::*;
412    /// use umbral::db::{RouteContext, TenantKey};
413    ///
414    /// App::builder()
415    ///     .route_context(|req| match req.headers().get("x-tenant") {
416    ///         Some(v) => RouteContext::new()
417    ///             .with_tenant(TenantKey::new(v.to_str().unwrap_or_default())),
418    ///         None => RouteContext::new(),
419    ///     })
420    ///     .build()?;
421    /// ```
422    pub fn route_context<F>(mut self, resolver: F) -> Self
423    where
424        F: Fn(&crate::web::Request) -> crate::db::RouteContext + Send + Sync + 'static,
425    {
426        self.route_context_resolver = Some(std::sync::Arc::new(resolver));
427        self
428    }
429
430    /// Register every `#[derive(Model)]` type the binary links, instead of naming
431    /// each one (gaps3 #46).
432    ///
433    /// ```ignore
434    /// App::builder()
435    ///     .settings(settings)
436    ///     .database("default", pool)
437    ///     .auto_models()          // replaces .model::<Post>().model::<Tag>()....
438    ///     .plugin(RestPlugin::default())
439    ///     .build()?
440    /// ```
441    ///
442    /// Explicit `.model::<T>()` still works and composes with this — the two are
443    /// merged and de-duplicated by table, so adding it to an existing app is safe.
444    ///
445    /// # Why this is opt-in
446    ///
447    /// Discovery is link-time. A model in your **binary** crate is always linked
448    /// and always found. A model in a **library** crate that nothing else
449    /// references can be dropped by the linker, and it would then be missing from
450    /// the registry — which means missing from `makemigrations`, i.e. a table that
451    /// silently never gets created. Making this the default would trade a little
452    /// typing for a failure mode that is invisible until production.
453    ///
454    /// If your models live in a library crate, either `use` it from `main.rs` (so
455    /// the linker keeps it) or keep naming them with `.model::<T>()`.
456    pub fn auto_models(mut self) -> Self {
457        self.auto_models = true;
458        self
459    }
460
461    /// Register a model with the app's migration engine.
462    ///
463    /// Called once per model the user wants the M5 `makemigrations` /
464    /// `migrate` commands to track. Captures the model's `NAME` /
465    /// `TABLE` / `FIELDS` constants into an owned `ModelMeta` so the
466    /// migration code can iterate without naming concrete `T` at the
467    /// call site. M7's Plugin contract will replace this with
468    /// `Plugin::models()` discovered through the plugin registry.
469    pub fn model<T: Model>(mut self) -> Self {
470        self.models.push(ModelMeta::for_::<T>());
471        self
472    }
473
474    /// Register one project-owned management command.
475    ///
476    /// The command shows up in `cargo run -- <name>`, in `umbral help`, and
477    /// under `umbral <name> --help`, exactly like a plugin's. The difference
478    /// is ownership: this one belongs to the binary, so there is no plugin
479    /// to wrap it in and nothing to publish.
480    ///
481    /// ```ignore
482    /// use umbral::cli::{CliError, PluginCommand, clap};
483    ///
484    /// struct BackfillSlugs;
485    ///
486    /// #[umbral::async_trait]
487    /// impl PluginCommand for BackfillSlugs {
488    ///     fn command(&self) -> clap::Command {
489    ///         clap::Command::new("backfill_slugs").about("Fill empty post slugs")
490    ///     }
491    ///     async fn run(&self, _m: &clap::ArgMatches) -> Result<(), CliError> {
492    ///         Ok(())
493    ///     }
494    /// }
495    ///
496    /// App::builder().command(BackfillSlugs)
497    /// ```
498    ///
499    /// `umbral startcommand` writes that file for you and wires this call.
500    ///
501    /// On a name clash with a plugin's command, the app's wins — the
502    /// project is the most specific layer — and the losing plugin is named
503    /// in a warning.
504    ///
505    /// A framework built-in (`migrate`, `serve`, …) cannot be overridden: the
506    /// dispatcher drops any registered command that lands on one of those
507    /// names and prints a warning telling you to rename it. That is enforced
508    /// in `cli::collect_commands`, not merely checked by `startcommand` —
509    /// a command named `migrate` would otherwise quietly take over, and the
510    /// next deploy would apply zero migrations and exit 0.
511    pub fn command(mut self, command: impl crate::cli::PluginCommand) -> Self {
512        self.commands.push(Box::new(command));
513        self
514    }
515
516    /// Register a whole list of project-owned commands at once — what the
517    /// generated `src/commands/mod.rs` hands back from its `all()` registry.
518    ///
519    /// ```ignore
520    /// App::builder().commands(commands::all())
521    /// ```
522    ///
523    /// That indirection is what makes `startcommand` idempotent: adding a
524    /// second command appends to `all()` and never touches `main.rs` again.
525    /// Rust has no way to discover a module by scanning a directory at
526    /// runtime, so `all()` *is* the auto-detection — a registry the tool
527    /// maintains for you.
528    pub fn commands(mut self, commands: Vec<Box<dyn crate::cli::PluginCommand>>) -> Self {
529        self.commands.extend(commands);
530        self
531    }
532
533    /// Register a plugin (M7).
534    ///
535    /// Plugins contribute models, routes, system_checks, and an
536    /// `on_ready` hook. `App::build()` topologically sorts the
537    /// registered set by `Plugin::dependencies()` and walks every
538    /// plugin's contributions. The plugin name `"app"` is reserved
539    /// for the implicit plugin that owns models registered via
540    /// `.model::<T>()`; a plugin claiming that name causes
541    /// `BuildError::ReservedPluginName`.
542    pub fn plugin<P: Plugin>(mut self, plugin: P) -> Self {
543        self.plugins.push(Box::new(plugin));
544        self
545    }
546
547    /// Attach a [`Routes`](crate::routes::Routes) bundle of
548    /// hand-registered routes.
549    ///
550    /// Each `.get(...) / .post(...) / .put(...) / .patch(...) /
551    /// .delete(...) / .head(...) / .options(...)` call on `Routes`
552    /// records the path *and* registers the handler, so the framework
553    /// surfaces declared routes in the dev-mode 404 page without a
554    /// parallel declaration list.
555    ///
556    /// Multi-method routes go through [`Routes::route`] (explicit
557    /// method list + `axum::routing::MethodRouter`). Routes that need
558    /// axum features the per-method shorthands don't expose (typed
559    /// `State`, middleware layers, `nest`, fallback handlers, etc.)
560    /// go through [`Routes::with_router`] — that escape hatch merges
561    /// an external `axum::Router` and its paths stay opaque to the
562    /// framework (won't appear in the dev 404 page).
563    ///
564    /// Calling this more than once merges the router and concatenates
565    /// the specs.
566    ///
567    /// ```ignore
568    /// use umbral::prelude::*;
569    ///
570    /// App::builder()
571    ///     .routes(
572    ///         Routes::new()
573    ///             .get("/", home)
574    ///             .get("/articles", list_articles_html)
575    ///             .post("/api/articles", create_article),
576    ///     )
577    ///     .build()?;
578    /// ```
579    pub fn routes(mut self, routes: crate::routes::Routes) -> Self {
580        let (router, specs) = routes.into_parts();
581        self.router = Some(match self.router.take() {
582            Some(prior) => prior.merge(router),
583            None => router,
584        });
585        self.route_paths.extend(specs);
586        self
587    }
588
589    /// Set the project-level templates directory.
590    ///
591    /// Defaults to `./templates` (relative to the binary's cwd) when
592    /// the builder method isn't called. If the resolved path doesn't
593    /// exist, the engine still publishes — calls to
594    /// `umbral::templates::render` then return `TemplateError::Missing`
595    /// with a clear diagnostic, which matches the "absence isn't an
596    /// error unless something tries to render" rule from the spec.
597    ///
598    /// This directory is searched first (highest priority). Plugin
599    /// directories contributed via `Plugin::templates_dirs()` are
600    /// appended in topological order and searched afterwards. To
601    /// override a plugin's template, drop a same-named file here.
602    pub fn templates_dir<P: Into<std::path::PathBuf>>(mut self, path: P) -> Self {
603        self.templates_dir = Some(path.into());
604        self
605    }
606
607    /// Set the trailing-slash redirect policy. See
608    /// [`crate::slash::SlashRedirect`].
609    ///
610    /// Default is `Off` (axum's strict matching). Most apps want
611    /// `Append` (`/foo` 404 → 308 → `/foo/`) so that
612    /// the same URL works with or without the trailing slash.
613    ///
614    /// ```ignore
615    /// use umbral::prelude::*;
616    /// use umbral::web::SlashRedirect;
617    ///
618    /// App::builder()
619    ///     .slash_redirect(SlashRedirect::Append)
620    ///     .build()?;
621    /// ```
622    pub fn slash_redirect(mut self, policy: crate::slash::SlashRedirect) -> Self {
623        self.slash_redirect = policy;
624        self
625    }
626
627    /// Set the template rendered on a 404. Follows the
628    /// `404.html` convention.
629    ///
630    /// The template gets `{ path }` in scope — the request path that
631    /// missed — so you can render `The page {{ path }} doesn't
632    /// exist.` without wiring extractors. When unset, 404s return
633    /// plain-text "Not Found". When set but the template fails to
634    /// render (missing file, parse error), the framework falls back
635    /// to the plain-text response and logs the render error.
636    ///
637    /// Composes with [`Self::slash_redirect`] — if a slash-redirect
638    /// probe finds the alternate, it 308s before the not-found
639    /// template fires.
640    pub fn not_found_template(mut self, name: impl Into<String>) -> Self {
641        self.not_found_template = Some(name.into());
642        self
643    }
644
645    /// Set the template rendered on a panicking handler. Follows
646    /// the `500.html` convention.
647    ///
648    /// Installs a `tower-http` `CatchPanic` layer around the router.
649    /// A panic in any handler is caught, logged via `tracing::error`,
650    /// and replaced with a 500 response carrying the rendered
651    /// template. When unset, panics use tower-http's default
652    /// behaviour (log + empty 500 body).
653    ///
654    /// In dev mode (`settings.environment == Dev`), the template receives
655    /// `dev_mode`, `error_display`, `error_chain`, and `request_path`
656    /// context variables. In prod those variables are empty.
657    ///
658    /// See [`Self::on_server_error`] for a hook that fires before the
659    /// template renders.
660    pub fn server_error_template(mut self, name: impl Into<String>) -> Self {
661        self.server_error_template = Some(name.into());
662        self
663    }
664
665    /// Register a custom template for error responses with `status` (e.g.
666    /// `429`, `403`, `410`). When a handler returns `Err((status, message))`
667    /// (or any non-HTML error response with this status), the template is
668    /// rendered in its place — styled like the 404/500 pages — preserving the
669    /// status code. The template receives `{ status, status_text, message,
670    /// request_path, dev_mode }`. Repeatable for multiple codes.
671    ///
672    /// 404 and 500 have dedicated methods ([`Self::not_found_template`] /
673    /// [`Self::server_error_template`]); use this for everything else.
674    ///
675    /// ```ignore
676    /// App::builder()
677    ///     .error_template(StatusCode::TOO_MANY_REQUESTS, "errors/429.html")
678    ///     .error_template(StatusCode::FORBIDDEN, "errors/403.html")
679    /// ```
680    pub fn error_template(
681        mut self,
682        status: axum::http::StatusCode,
683        name: impl Into<String>,
684    ) -> Self {
685        self.error_templates.insert(status, name.into());
686        self
687    }
688
689    /// Register a hook that fires on every internal server error (500).
690    ///
691    /// The closure receives:
692    /// - `error_display: &str` — the `Display` form of the error or the
693    ///   stringified panic payload.
694    /// - `request_path: &str` — the URI path of the failing request (empty
695    ///   for panic-path errors where path isn't yet available).
696    ///
697    /// The hook runs synchronously before the 500 template is rendered. It
698    /// cannot change the response — use it to log to an external service
699    /// (Sentry, Datadog, a file, etc.).
700    ///
701    /// ```ignore
702    /// App::builder()
703    ///     .on_server_error(|err, path| {
704    ///         tracing::error!(err, path, "500 error");
705    ///     })
706    ///     .build()?
707    /// ```
708    pub fn on_server_error<F>(mut self, hook: F) -> Self
709    where
710        F: Fn(&str, &str) + Send + Sync + 'static,
711    {
712        self.server_error_hook = Some(std::sync::Arc::new(hook));
713        self
714    }
715
716    /// Disable the built-in default 404/500 templates.
717    ///
718    /// By default, when the user hasn't called `.not_found_template(...)` or
719    /// `.server_error_template(...)`, umbral renders its own embedded Tailwind
720    /// error pages. Call this method to revert to axum's built-in behaviour:
721    /// a plain-text "Not Found" on 404 and an empty 500 body on panic.
722    ///
723    /// ```ignore
724    /// App::builder()
725    ///     .disable_default_error_pages()
726    ///     .build()?
727    /// ```
728    /// gaps3 #23: apply pending migrations automatically when the app is
729    /// STARTED (`umbral_cli::dispatch` → the `serve` command), and NEVER during
730    /// `makemigrations` / `migrate` / any other subcommand. This replaces the
731    /// argv-sniffing guard consumers hand-rolled in `main.rs` to avoid
732    /// auto-migrating during CLI commands:
733    ///
734    /// ```ignore
735    /// let app = App::builder().auto_migrate_on_serve().plugin(...).build()?;
736    /// umbral_cli::dispatch(app).await   // migrate runs iff this serves
737    /// ```
738    ///
739    /// A convenience for demos / small apps; a large deploy still runs
740    /// `migrate` as an explicit release step. Seeding stays app-owned (a
741    /// plugin's `on_ready` or an explicit call).
742    pub fn auto_migrate_on_serve(mut self) -> Self {
743        self.auto_migrate_on_serve = true;
744        self
745    }
746
747    /// Drain for `delay` on shutdown before the server stops accepting
748    /// connections (Kikosi #5 — the zero-downtime rollout piece).
749    ///
750    /// On `SIGTERM` / Ctrl-C, [`App::serve`] marks the process draining
751    /// ([`crate::shutdown::is_draining`]) so `umbral-health`'s `/readyz` returns
752    /// 503 at once, keeps serving for `delay`, and only then lets the graceful
753    /// shutdown proceed (stop accepting, finish in-flight, close pools). The
754    /// delay is the window in which a load balancer polls `/readyz`, sees the
755    /// 503, and pulls this instance out of rotation — so the requests it *would*
756    /// have routed here go elsewhere instead of hitting a socket that is about
757    /// to close.
758    ///
759    /// Pick a delay a little longer than your LB's readiness probe interval
760    /// (k8s default 10s; a `HEALTHCHECK --interval=10s` the same). `5`–`15s` is
761    /// typical. `Duration::ZERO` (the default) skips the drain entirely — right
762    /// for a single-instance app or local dev, where there is no LB to notify
763    /// and an instant Ctrl-C is what you want.
764    ///
765    /// Only meaningful alongside a readiness probe (mount `HealthPlugin`); with
766    /// no `/readyz` for the LB to poll, the delay is dead time on shutdown.
767    pub fn shutdown_drain(mut self, delay: std::time::Duration) -> Self {
768        self.drain_delay = delay;
769        self
770    }
771
772    pub fn disable_default_error_pages(mut self) -> Self {
773        self.default_error_pages = false;
774        self
775    }
776
777    /// Install a CORS policy as the outermost middleware.
778    ///
779    /// The framework doesn't install a `CorsLayer` by default —
780    /// same-origin requests need no policy, and CORS is too
781    /// security-sensitive to enable implicitly. Pass a
782    /// [`crate::cors::CorsConfig`] (start from
783    /// [`CorsConfig::strict`](crate::cors::CorsConfig::strict) for
784    /// production or [`CorsConfig::permissive`](crate::cors::CorsConfig::permissive)
785    /// for dev).
786    ///
787    /// ```ignore
788    /// use umbral::prelude::*;
789    /// use umbral::cors::CorsConfig;
790    ///
791    /// App::builder()
792    ///     .cors(CorsConfig::strict()
793    ///         .allow_origin("https://app.example.com")
794    ///         .allow_credentials(true))
795    ///     .build()
796    ///     .await?
797    /// ```
798    ///
799    /// The layer is applied LAST in the middleware chain so it
800    /// becomes the outermost wrapper — preflight `OPTIONS` is
801    /// answered before any plugin / handler sees the request, and
802    /// the response headers are added on the way back out
803    /// regardless of which downstream layer produced the body.
804    pub fn cors(mut self, config: crate::cors::CorsConfig) -> Self {
805        self.cors = Some(config);
806        self
807    }
808
809    /// Apply a CORS policy scoped to requests whose path starts with `prefix`
810    /// (e.g. `"/api"`), leaving every other route's responses untouched. The
811    /// path-scoped counterpart to [`cors`](Self::cors) — the shape you want for
812    /// "CORS on the REST API, not the HTML pages." Call repeatedly for several
813    /// prefixes. Scoped policies are applied after (outside) the global one.
814    ///
815    /// ```ignore
816    /// use umbral::cors::CorsConfig;
817    ///
818    /// App::builder()
819    ///     .cors_for("/api", CorsConfig::strict()
820    ///         .allow_origins(vec!["https://app.example.com"])
821    ///         .allow_credentials(true))
822    ///     .build()
823    ///     .await?
824    /// ```
825    pub fn cors_for(mut self, prefix: impl Into<String>, config: crate::cors::CorsConfig) -> Self {
826        self.cors_scoped.push((prefix.into(), config));
827        self
828    }
829
830    /// Default every ORM write to run inside its own transaction.
831    ///
832    /// When `enabled = true`, terminals that opt into the contract
833    /// (`Manager::create`, `Manager::bulk_create`,
834    /// `Manager::get_or_create`, `QuerySet::update_values`,
835    /// `QuerySet::delete`) wrap their work in a BEGIN / COMMIT pair
836    /// unless the caller explicitly opts out with `.non_atomic()`.
837    ///
838    /// This is the safe-by-default posture: a framework that claims
839    /// "secure by default" should also be "transaction-safe by
840    /// default." Opting out matters mostly for high-throughput seed
841    /// scripts that already wrap an outer transaction themselves.
842    ///
843    /// Without this flag the framework's behaviour is unchanged —
844    /// writes run with whatever transaction the caller arranges. The
845    /// per-call `.atomic()` / `.non_atomic()` overrides still work.
846    pub fn atomic_transactions(mut self, enabled: bool) -> Self {
847        self.atomic_transactions = Some(enabled);
848        self
849    }
850
851    /// Make a forgotten authorization gate a **boot error** instead of a
852    /// warning (gaps3 #28 P1). With this set, `build()` fails with
853    /// [`BuildError::UngatedMutatingRoutes`] if any app-level mutating route
854    /// (POST/PUT/PATCH/DELETE) registered via [`Self::routes`] carries no
855    /// recorded permission — i.e. it wasn't gated through the umbral-permissions
856    /// `Routes::*_gated(...)` builders (a hand-applied
857    /// `.layer(permission_required(...))` is opaque to the audit, so prefer the
858    /// builder). This is the opt-in "gated by construction" posture: authorization
859    /// on every mutating route is enforced at boot rather than trusted to review.
860    ///
861    /// Default off — `build()` only *warns*. An intentionally-public mutating
862    /// route (a webhook receiver, a health `POST`) must be registered through a
863    /// permission-aware builder anyway (or kept out of `.routes(...)`) once this
864    /// is on, so the decision is explicit.
865    pub fn deny_ungated_mutations(mut self) -> Self {
866        self.deny_ungated_mutations = true;
867        self
868    }
869
870    /// Compress responses with gzip / brotli (a `tower-http`
871    /// `CompressionLayer`). The algorithm is chosen from the request's
872    /// `Accept-Encoding`; already-encoded or non-compressible content types
873    /// are skipped automatically.
874    ///
875    /// Off by default: in most deployments the reverse proxy (nginx, a CDN)
876    /// already compresses, and doing it twice is wasted CPU. Enable this
877    /// when you serve directly (a single binary with no proxy in front).
878    pub fn compression(mut self) -> Self {
879        self.compress = true;
880        self
881    }
882
883    /// Set (or disable) the framework-wide request-body size cap.
884    ///
885    /// `build()` installs a `tower-http` `RequestBodyLimitLayer` with this
886    /// ceiling, so any request whose body exceeds it is rejected with `413
887    /// Payload Too Large` before a handler (or the multipart parser) buffers
888    /// it — the memory-exhaustion backstop axum's per-extractor default does
889    /// NOT give streaming/multipart consumers (audit_2 core-web H11).
890    ///
891    /// Defaults to **32 MiB**. Pass `Some(bytes)` to raise/lower it, or `None`
892    /// to remove the global limit entirely (appropriate when a reverse proxy
893    /// already caps body size).
894    ///
895    /// ```ignore
896    /// App::builder()
897    ///     .max_request_body(Some(8 * 1024 * 1024)) // 8 MiB
898    ///     .build().await?;
899    /// ```
900    pub fn max_request_body(mut self, limit: Option<usize>) -> Self {
901        self.max_request_body_bytes = limit;
902        self
903    }
904
905    /// Set (or disable) the default per-request timeout.
906    ///
907    /// `build()` installs a `tower-http` `TimeoutLayer` so a request that runs
908    /// longer than this is aborted with `408 Request Timeout`, freeing the
909    /// task/connection instead of letting a hung handler or slowloris client
910    /// pin it indefinitely (audit_2 core-web H11/#3).
911    ///
912    /// Defaults to **30 seconds**. Pass `Some(duration)` to change it, or
913    /// `None` to disable — do that for legitimately long-lived streaming/SSE
914    /// routes, or when a proxy owns request timeouts.
915    ///
916    /// ```ignore
917    /// use std::time::Duration;
918    /// App::builder()
919    ///     .request_timeout(Some(Duration::from_secs(10)))
920    ///     .build().await?;
921    /// ```
922    pub fn request_timeout(mut self, timeout: Option<std::time::Duration>) -> Self {
923        self.request_timeout = timeout;
924        self
925    }
926
927    /// Toggle the core-shipped hardening response headers (audit_2 H10):
928    /// `X-Content-Type-Options: nosniff`, `X-Frame-Options: DENY`, and
929    /// `Referrer-Policy: strict-origin-when-cross-origin`. On by default and
930    /// applied only when the header isn't already set, so `SecurityPlugin`'s
931    /// configured values win. Pass `false` to fully own response headers
932    /// yourself (e.g. an API behind a gateway that adds them at the edge).
933    pub fn default_security_headers(mut self, enabled: bool) -> Self {
934        self.default_security_headers = enabled;
935        self
936    }
937
938    /// Register a framework-level [`Middleware`](crate::middleware::Middleware)
939    /// (feature #68) with `before_request` / `after_response` hooks.
940    ///
941    /// App-level middleware is added to the stack *before* any plugin's
942    /// contribution, so its `before_request` runs first and its
943    /// `after_response` runs last (it's the outermost layer of the onion).
944    /// Call this multiple times to register several, in order.
945    ///
946    /// Use this for the common "look at every request / response" case.
947    /// For a real tower `Layer` (timeouts, body limits) reach for the
948    /// router directly via a plugin's `wrap_router`.
949    pub fn middleware<M: crate::middleware::Middleware>(mut self, mw: M) -> Self {
950        self.middleware.push(std::sync::Arc::new(mw));
951        self
952    }
953
954    /// Finalize the application.
955    ///
956    /// Phases (see spec 01 §Mechanics and invariants and spec 02
957    /// §Dependency ordering):
958    ///
959    /// 1. **Collect.** Gather settings, databases, and router from
960    ///    builder-local state. Settings must be set explicitly via
961    ///    `.settings(...)`; the "default" database pool must be
962    ///    registered via `.database("default", pool)`. The caller
963    ///    opens the pool first (with `umbral::db::connect(...).await`)
964    ///    and hands it to the builder. This matches the canonical
965    ///    pattern in spec 01-app-and-settings.md.
966    /// 2. **Validate plugins.** Reject the reserved `"app"` name,
967    ///    reject duplicate `Plugin::name()`s, verify every entry in a
968    ///    `dependencies()` list points at a registered plugin, and
969    ///    compute a stable topological order. Cycles surface as
970    ///    `BuildError::PluginCycle`.
971    /// 3. **Detect backend.** `backend::detect(&settings.database_url)`
972    ///    picks one of the shipped `DatabaseBackend` impls (M4
973    ///    abstraction). An unknown URL scheme (mysql / oracle / etc.)
974    ///    fails here, before any system check runs.
975    /// 4. **Publish ambient state.** Write settings, pools, and the
976    ///    active backend into their `OnceLock`s. The model registry
977    ///    carries one entry per plugin (the implicit `"app"` plus every
978    ///    registered plugin's `Plugin::models()`).
979    /// 5. **System check.** Run framework-built-in checks plus every
980    ///    plugin's `system_checks()` (concatenated in topological order)
981    ///    against the just-published context. Errors block boot;
982    ///    warnings log and continue.
983    /// 6. **Build router.** Start from the hand-written router (or a
984    ///    fallback handler), then merge every plugin's `routes()` in
985    ///    topological order. axum's `Router::merge` panics on
986    ///    duplicate routes with a clear message.
987    /// 7. **Fire `on_ready`.** Call each plugin's `on_ready(&AppContext)`
988    ///    in topological order. A failure here surfaces as
989    ///    `BuildError::PluginOnReady`. Phases 1-6 are
990    ///    [`AppBuilder::build_deferred`]; this last phase is [`App::ready`],
991    ///    and a CLI binary lets `umbral_cli::dispatch` decide when it fires
992    ///    (gaps3 #41).
993    ///
994    /// `build()` is intentionally sync. Earlier iterations auto-opened
995    /// the default pool from `settings.database_url` by spinning up a
996    /// throwaway tokio runtime to drive `db::connect`. That panicked
997    /// when called from inside any caller that was already in a tokio
998    /// runtime ("Cannot start a runtime from within a runtime"), which
999    /// is every realistic case. Requiring an explicit `.database(...)`
1000    /// is both spec-correct and avoids the trap.
1001    pub fn build(self) -> Result<App, BuildError> {
1002        let app = self.build_deferred()?;
1003        app.ready()?;
1004        Ok(app)
1005    }
1006
1007    /// Everything [`AppBuilder::build`] does *except* firing `on_ready`.
1008    ///
1009    /// The app is fully wired — pools open, registry published, router merged,
1010    /// system checks passed — but no plugin has been told the app is up. Call
1011    /// [`App::ready`] when it actually is.
1012    ///
1013    /// This exists for `umbral_cli::dispatch`, which has to build the app in
1014    /// order to read its plugins' `commands()` and *then* decide what argv asked
1015    /// for. A schema command (`migrate`, `makemigrations`, `inspectdb`) must not
1016    /// fire hooks that seed content into a schema that doesn't exist yet
1017    /// (gaps3 #41). Reach for it directly only if you are writing your own
1018    /// dispatcher; otherwise `build()` is the one you want.
1019    pub fn build_deferred(mut self) -> Result<App, BuildError> {
1020        // Phase 1 — collect
1021        let settings = self.settings.take().ok_or(BuildError::SettingsMissing)?;
1022
1023        if !self.databases.contains_key("default") {
1024            return Err(BuildError::DefaultPoolMissing);
1025        }
1026
1027        // Phase 1.4 — audit_2 H17: open the pools declared in `settings.databases`.
1028        // Each `[databases] <alias> = "<url>"` entry that a builder `.database()`
1029        // call didn't already register is opened LAZILY (sync; connects on first
1030        // use) and added to the pool set, so a model/router routed to that alias
1031        // resolves instead of panicking at query time — and the documented
1032        // `settings.databases` config actually does something. A builder-registered
1033        // alias wins (an explicitly-built pool overrides the settings URL).
1034        for (alias, url) in &settings.databases {
1035            if self.databases.contains_key(alias) {
1036                continue;
1037            }
1038            let pool =
1039                crate::db::connect_lazy(url).map_err(|error| BuildError::SettingsDatabasePool {
1040                    alias: alias.clone(),
1041                    error,
1042                })?;
1043            self.databases.insert(alias.clone(), pool);
1044        }
1045
1046        // Phase 1.5 — validate plugins and compute a stable topological
1047        // order. Reserved-name and duplicate-name checks reject the
1048        // build before any ambient state gets published; the toposort
1049        // surfaces both missing deps and cycles as `BuildError`. The
1050        // sorted vec is reused in phases 3 / 4 / 5 / 6 so every plugin
1051        // walk reads from one canonical order, then handed to `App` so
1052        // post-build callers (notably `umbral::cli::dispatch`) can walk
1053        // the same list.
1054        let sorted_plugins = sort_plugins(std::mem::take(&mut self.plugins))?;
1055
1056        // Phase 2 — detect backend from the configured URL.
1057        let backend =
1058            crate::backend::detect(&settings.database_url).map_err(BuildError::BackendDetect)?;
1059
1060        // Phase 2.1 — cross-check the registered default pool's
1061        // backend against the URL-derived one. A mismatch (e.g. the
1062        // URL says `sqlite://` but the caller passed in a `PgPool`)
1063        // surfaces here with a clear name pair rather than as a
1064        // confusing query-time error.
1065        let default_pool = self
1066            .databases
1067            .get("default")
1068            .expect("contains_key check above");
1069        if default_pool.backend_name() != backend.name() {
1070            return Err(BuildError::DatabaseBackendMismatch {
1071                url_backend: backend.name(),
1072                pool_backend: default_pool.backend_name(),
1073            });
1074        }
1075
1076        // Phase 2.5 — validate every plugin's `database()` alias
1077        // against the registered pool set BEFORE phase 3 moves
1078        // `self.databases` into the ambient registry. Lets a typo
1079        // surface at boot with a clear diagnostic instead of as a
1080        // runtime "no pool registered" panic from `db::pool_for`.
1081        // Also collect the per-model alias map for `init_model_aliases`
1082        // below. Two layers: plugin-level (`Plugin::database()`) and
1083        // per-model (`#[umbral(database = "alias")]` → `Model::DATABASE`,
1084        // surfaced via `ModelMeta::database`). Per-model wins when both
1085        // are set — useful for a plugin that owns one model on the
1086        // primary DB and another on an analytics/archive DB. Same alias
1087        // validation: a typo surfaces at boot, not at runtime.
1088        let mut model_aliases: HashMap<String, String> = HashMap::new();
1089        for plugin in &sorted_plugins {
1090            // Plugin-level default for every model this plugin contributes.
1091            if let Some(alias) = plugin.database() {
1092                if !self.databases.contains_key(alias) {
1093                    return Err(BuildError::PluginDatabaseAlias {
1094                        plugin: plugin.name(),
1095                        alias,
1096                    });
1097                }
1098                for model in plugin.models() {
1099                    model_aliases.insert(model.name, alias.to_string());
1100                }
1101            }
1102            // Per-model overrides — walked AFTER the plugin pass so they
1103            // can supersede the plugin's choice.
1104            for model in plugin.models() {
1105                if let Some(alias) = &model.database {
1106                    if !self.databases.contains_key(alias) {
1107                        return Err(BuildError::PluginDatabaseAlias {
1108                            plugin: plugin.name(),
1109                            alias: Box::leak(alias.clone().into_boxed_str()),
1110                        });
1111                    }
1112                    model_aliases.insert(model.name.clone(), alias.clone());
1113                }
1114            }
1115        }
1116        // Same per-model walk for the implicit `"app"` plugin's
1117        // user-registered models, which don't have a `Plugin::database()`
1118        // wrapper to inherit from.
1119        for model in &self.models {
1120            if let Some(alias) = &model.database {
1121                if !self.databases.contains_key(alias) {
1122                    return Err(BuildError::PluginDatabaseAlias {
1123                        plugin: crate::migrate::APP_PLUGIN_NAME,
1124                        alias: Box::leak(alias.clone().into_boxed_str()),
1125                    });
1126                }
1127                model_aliases.insert(model.name.clone(), alias.clone());
1128            }
1129        }
1130
1131        // (audit_2 H17: `settings.databases` pools were opened lazily in Phase 1.4
1132        // above, so every declared alias is now a registered pool — the earlier
1133        // "not auto-opened" boot warning is gone.)
1134
1135        // Phase 2.5b — cross-database foreign-key guard (gaps2 #22).
1136        //
1137        // A foreign key whose target model lives on a DIFFERENT database
1138        // can't be a real DB constraint — `REFERENCES` can't span pools.
1139        // We resolve each model's effective alias (plugin default, then
1140        // per-model override, else "default") into a table→alias map,
1141        // then check every FK column: if the column's target table
1142        // routes to a different alias than the model AND the field has
1143        // not opted out via `#[umbral(db_constraint = false)]`, the build
1144        // fails loudly here rather than emitting an invalid `FOREIGN KEY`
1145        // line at migration time.
1146        //
1147        // Build the table→alias map with the same precedence as
1148        // `model_aliases` above: plugin default first, per-model override
1149        // wins, the implicit "app" models last. Any table not mentioned
1150        // routes to "default".
1151        let mut table_alias: HashMap<String, String> = HashMap::new();
1152        for plugin in &sorted_plugins {
1153            let plugin_default = plugin.database();
1154            for model in plugin.models() {
1155                let alias = model
1156                    .database
1157                    .clone()
1158                    .or_else(|| plugin_default.map(|s| s.to_string()))
1159                    .unwrap_or_else(|| "default".to_string());
1160                table_alias.insert(model.table.clone(), alias);
1161            }
1162        }
1163        for model in &self.models {
1164            let alias = model
1165                .database
1166                .clone()
1167                .unwrap_or_else(|| "default".to_string());
1168            table_alias.insert(model.table.clone(), alias);
1169        }
1170        // Helper to resolve a table's alias, defaulting to "default".
1171        let alias_of = |table: &str| -> String {
1172            table_alias
1173                .get(table)
1174                .cloned()
1175                .unwrap_or_else(|| "default".to_string())
1176        };
1177        // Walk every model's FK fields and check each FK relation. The
1178        // default (no custom router) path keeps today's build-time local
1179        // alias equality (`alias_of(a) == alias_of(b)`): the trait's
1180        // DEFAULT `allow_relation` reads the GLOBAL `model_alias`, which is
1181        // still unpublished at this Phase 2.5b point, so routing the
1182        // default case through the trait would compare "default" == "default"
1183        // for everything and silently disable the #22 guard. A CUSTOM router
1184        // is asked directly via `allow_relation`.
1185        //
1186        // gaps3 #46: pull in every model the binary link-registered. Merged with
1187        // (not instead of) the explicit `.model::<T>()` list, and de-duplicated by
1188        // table, so the two compose and adding `auto_models()` to an existing app
1189        // can't double-register anything.
1190        if self.auto_models {
1191            let known: std::collections::HashSet<String> = self
1192                .models
1193                .iter()
1194                .map(|m| m.table.clone())
1195                .chain(
1196                    sorted_plugins
1197                        .iter()
1198                        .flat_map(|p| p.models())
1199                        .map(|m| m.table),
1200                )
1201                .collect();
1202            for meta in crate::migrate::link_registered_models() {
1203                if !known.contains(&meta.table) {
1204                    self.models.push(meta);
1205                }
1206            }
1207        }
1208
1209        // gaps3 #54: an `#[umbral(audited)]` model implies the audit table. Register
1210        // it automatically so `makemigrations` creates it through the normal
1211        // declare→migrate loop — no special-cased DDL, and no ceremony for the app.
1212        let any_audited = sorted_plugins
1213            .iter()
1214            .flat_map(|p| p.models())
1215            .chain(self.models.iter().cloned())
1216            .any(|m| m.audited);
1217        if any_audited
1218            && !self
1219                .models
1220                .iter()
1221                .any(|m| m.table == crate::orm::audit::AUDIT_TABLE)
1222        {
1223            self.models.push(crate::orm::audit::audit_meta());
1224        }
1225
1226        // Materialize the models into a Vec so we can both build a
1227        // table→meta lookup AND iterate them.
1228        let all_models: Vec<ModelMeta> = sorted_plugins
1229            .iter()
1230            .flat_map(|p| p.models())
1231            .chain(self.models.iter().cloned())
1232            .collect();
1233        let meta_by_table: HashMap<&str, &ModelMeta> =
1234            all_models.iter().map(|m| (m.table.as_str(), m)).collect();
1235        // Clone the candidate router — install still happens at Phase 3, so
1236        // we must NOT take/consume `self.db_router` here.
1237        let candidate_router = self.db_router.clone();
1238        for model in &all_models {
1239            for field in &model.fields {
1240                let Some(target_table) = field.fk_target.as_deref() else {
1241                    continue;
1242                };
1243                if !field.db_constraint {
1244                    continue;
1245                }
1246                let allowed = match &candidate_router {
1247                    Some(r) => match meta_by_table.get(target_table) {
1248                        Some(target_meta) => r.allow_relation(model, target_meta),
1249                        // Target isn't a registered model (shouldn't happen
1250                        // for a real FK); don't false-reject — fall back to
1251                        // the local alias check.
1252                        None => alias_of(&model.table) == alias_of(target_table),
1253                    },
1254                    // No custom router: today's build-time local alias
1255                    // equality (#22).
1256                    None => alias_of(&model.table) == alias_of(target_table),
1257                };
1258                if !allowed {
1259                    let model_db = alias_of(&model.table);
1260                    let target_db = alias_of(target_table);
1261                    return Err(BuildError::CrossDatabaseForeignKey {
1262                        model: Box::leak(model.name.clone().into_boxed_str()),
1263                        field: Box::leak(field.name.clone().into_boxed_str()),
1264                        model_db: Box::leak(model_db.into_boxed_str()),
1265                        target_db: Box::leak(target_db.into_boxed_str()),
1266                    });
1267                }
1268            }
1269        }
1270
1271        // Phase 2.6 — publish the default-error-pages flag before the
1272        // templates engine starts so `errors::default_pages_enabled()` is
1273        // correct the moment any 404/500 helper is called.
1274        crate::errors::init_default_pages(self.default_error_pages);
1275
1276        // Phase 3 — publish ambient state. The model registry now carries
1277        // one entry per registered plugin (the implicit `"app"` plugin
1278        // for `.model::<T>()` registrations, plus every `.plugin(...)`
1279        // contribution). Plugins that contribute zero models still get a
1280        // map entry; the flattening in `migrate::init_plugins` collapses
1281        // them to nothing in the registry but the per-plugin model walk
1282        // stays deterministic.
1283        crate::settings::init(&settings);
1284        db::init(self.databases);
1285        if let Some(router) = self.db_router {
1286            crate::db::router::install_router(router);
1287        }
1288        crate::backend::init(backend);
1289        if let Some(enabled) = self.atomic_transactions {
1290            db::init_atomic_default(enabled);
1291        }
1292
1293        let mut per_plugin: HashMap<String, Vec<ModelMeta>> = HashMap::new();
1294        per_plugin.insert(
1295            crate::migrate::APP_PLUGIN_NAME.to_string(),
1296            std::mem::take(&mut self.models),
1297        );
1298        for plugin in &sorted_plugins {
1299            per_plugin.insert(plugin.name().to_string(), plugin.models());
1300        }
1301        crate::migrate::init_plugins(per_plugin);
1302
1303        // Publish the topological plugin order so the migration engine
1304        // walks plugins in dependency order. The implicit "app" plugin
1305        // (owner of `.model::<T>()` registrations) lands LAST: app models
1306        // typically hold ForeignKeys INTO plugin-owned tables (e.g.
1307        // `Post.author -> auth_user`), so those tables must be created
1308        // first. Postgres enforces FK targets at CREATE TABLE, so ordering
1309        // "app" first made app-model migrations fail there with
1310        // `relation "auth_user" does not exist` (SQLite silently allowed
1311        // the dangling FK, hiding the bug in local dev).
1312        let mut order: Vec<String> = Vec::with_capacity(sorted_plugins.len() + 1);
1313        for plugin in &sorted_plugins {
1314            order.push(plugin.name().to_string());
1315        }
1316        order.push(crate::migrate::APP_PLUGIN_NAME.to_string());
1317        crate::migrate::init_plugin_order(order);
1318
1319        // Collect every plugin's advertised API endpoints into a global
1320        // so a discovery surface (umbral-rest's API root) can list them
1321        // without depending on the contributing plugins' crates. In
1322        // registration order; plugins that advertise nothing contribute
1323        // nothing.
1324        let mut api_endpoints = Vec::new();
1325        for plugin in &sorted_plugins {
1326            api_endpoints.extend(plugin.api_endpoints());
1327        }
1328        crate::migrate::init_api_endpoints(api_endpoints);
1329
1330        // Publish the per-plugin model alias map collected in phase
1331        // 2.5. Done after `migrate::init_plugins` so the migration
1332        // registry is alive when QuerySet's resolve_pool starts
1333        // looking up by `Model::NAME`.
1334        crate::migrate::init_model_aliases(model_aliases);
1335
1336        // audit_2 H19: surface at boot the app's own mutating routes
1337        // (POST/PUT/PATCH/DELETE) that carry no RECORDED permission, so a
1338        // forgotten authorization gate surfaces here instead of as a silently
1339        // open endpoint. Only `.routes(...)` (the app's hand-written routes)
1340        // are audited — plugin routes gate via their own conventions and are
1341        // merged separately. Runs before `self.route_paths` is moved below.
1342        // With `.deny_ungated_mutations()` (gaps3 #28 P1) the same finding is a
1343        // hard `BuildError` instead of a warning: authorization on every
1344        // mutating route is enforced by construction.
1345        let ungated = ungated_mutating_routes(&self.route_paths);
1346        if !ungated.is_empty() {
1347            if self.deny_ungated_mutations {
1348                return Err(BuildError::UngatedMutatingRoutes { routes: ungated });
1349            }
1350            warn_ungated_mutating_routes(&ungated);
1351        }
1352
1353        // Snapshot the declared route paths into the registry so the
1354        // dev-mode 404 page can surface them. The implicit `"app"`
1355        // plugin holds whatever `.route_paths([...])` declared on the
1356        // builder; each registered plugin contributes its own list.
1357        // Empty entries are kept so the listing distinguishes "plugin
1358        // present, no routes" from "plugin absent".
1359        let mut route_registry = crate::routes::RouteRegistry::default();
1360        route_registry.by_plugin.insert(
1361            crate::migrate::APP_PLUGIN_NAME.to_string(),
1362            std::mem::take(&mut self.route_paths),
1363        );
1364        // gaps4 #31: when a plugin implements the recording `routes_builder()`,
1365        // its router AND its declared specs come from that ONE source, so the
1366        // registry cannot drift from what's mounted. Call it once here, record
1367        // the specs now, and stash the router for the merge loop below (so a
1368        // legacy plugin's `routes()` side-effects still fire at their original
1369        // point, and a builder plugin's router is never rebuilt twice).
1370        let mut builder_routers: HashMap<String, Router> = HashMap::new();
1371        for plugin in &sorted_plugins {
1372            let specs = match plugin.routes_builder() {
1373                Some(builder) => {
1374                    let (router, specs) = builder.into_parts();
1375                    builder_routers.insert(plugin.name().to_string(), router);
1376                    specs
1377                }
1378                None => plugin.route_paths(),
1379            };
1380            route_registry
1381                .by_plugin
1382                .insert(plugin.name().to_string(), specs);
1383        }
1384        crate::routes::init(route_registry);
1385
1386        // BUG-20: publish every plugin's OpenAPI path contribution
1387        // so umbral-openapi can merge them into the emitted spec.
1388        // Flat (path, value) list — multiple plugins contributing
1389        // the same path produce duplicate entries; umbral-openapi's
1390        // merge step picks the first.
1391        let mut openapi_entries: Vec<(String, serde_json::Value)> = Vec::new();
1392        for plugin in &sorted_plugins {
1393            openapi_entries.extend(plugin.openapi_paths());
1394        }
1395        crate::routes::init_openapi(openapi_entries);
1396
1397        // Templates engine — published before phase 4 so a future
1398        // plugin system_check that wants to inspect the loaded
1399        // templates can.
1400        //
1401        // Search order (first-match-wins across all template directories):
1402        //   1. App-level dir: set via `.templates_dir(...)` or `./templates`.
1403        //   2. Plugin dirs: each plugin's `templates_dirs()` contributions,
1404        //      in topological dependency order.
1405        //
1406        // The engine warns (via tracing) when two directories ship a
1407        // template with the same name — the first-registered copy wins.
1408        let app_templates_dir = self
1409            .templates_dir
1410            .take()
1411            .unwrap_or_else(|| std::path::PathBuf::from("templates"));
1412        let mut all_template_dirs: Vec<std::path::PathBuf> = vec![app_templates_dir];
1413        for plugin in &sorted_plugins {
1414            all_template_dirs.extend(plugin.templates_dirs());
1415        }
1416        // features.md #67 — collect every plugin's custom tags/filters in
1417        // topological order so a dependency's registrar runs before its
1418        // dependent's (and a later plugin can override an earlier one).
1419        let mut template_registrars: Vec<crate::templates::TemplateRegistrar> = Vec::new();
1420        for plugin in &sorted_plugins {
1421            template_registrars.extend(plugin.template_registrars());
1422        }
1423        // `init_with` returns the list of collision names (templates present
1424        // in more than one directory). We log each one via tracing here so
1425        // the `App::build()` phase is the single point that handles warnings;
1426        // `templates::init` itself also emits tracing::warn! for each, but
1427        // returning the list lets callers (tests) assert without a subscriber.
1428        let _collisions = crate::templates::init_with(&all_template_dirs, template_registrars)
1429            .map_err(BuildError::TemplatesInit)?;
1430
1431        // Phase 4 — system check. Build the context against ambient
1432        // state, run the framework checks plus every plugin's
1433        // contribution in topological order, partition into errors vs
1434        // warnings, log the warnings, fail the build on any errors.
1435        // Whether any registered plugin declares a Storage backend. Read
1436        // by the `field.storage_backend` check; computed from the
1437        // capability flag (not the ambient `storage_opt()`) because
1438        // backends register in `on_ready`, which runs *after* this phase.
1439        let provides_storage = sorted_plugins.iter().any(|p| p.provides_storage());
1440        let plugin_names: Vec<&str> = sorted_plugins.iter().map(|p| p.name()).collect();
1441        let ctx = crate::check::CheckContext {
1442            backend,
1443            settings: crate::settings::get(),
1444            provides_storage,
1445            registered_plugin_names: &plugin_names,
1446        };
1447        let mut checks = crate::check::framework_checks();
1448        for plugin in &sorted_plugins {
1449            checks.extend(plugin.system_checks());
1450        }
1451        let findings = crate::check::run_all(&ctx, &checks);
1452        let mut errors = Vec::new();
1453        for finding in findings {
1454            match finding.severity {
1455                crate::check::Severity::Error => errors.push(finding),
1456                crate::check::Severity::Warning => {
1457                    tracing::warn!(
1458                        check = finding.check_id,
1459                        "umbral system check warning: {}",
1460                        finding.message
1461                    );
1462                }
1463            }
1464        }
1465        if !errors.is_empty() {
1466            return Err(BuildError::SystemCheckFailed { findings: errors });
1467        }
1468
1469        // Phase 5 — build the merged router. Start from the hand-written
1470        // router (or a fallback handler if none was registered), then
1471        // merge every plugin's routes in topological order. axum's
1472        // `Router::merge` composes path tables; conflicts panic with a
1473        // clear message.
1474        let mut router = self.router.unwrap_or_else(|| {
1475            Router::new().fallback(|| async { "umbral is running, but no routes are registered." })
1476        });
1477        for plugin in &sorted_plugins {
1478            // gaps4 #31: a plugin that provided a `routes_builder()` had its
1479            // router built above (drift-free with its specs); reuse it. Everyone
1480            // else mounts through the legacy `routes()`, unchanged.
1481            let plugin_router = builder_routers
1482                .remove(plugin.name())
1483                .unwrap_or_else(|| plugin.routes());
1484            router = router.merge(plugin_router);
1485            // Phase 5.4 — mount the plugin's `include_bytes!`-embedded
1486            // assets. Each StaticFile becomes a GET route serving the
1487            // body with the supplied content-type + cache-control.
1488            for file in plugin.static_files() {
1489                router = router.route(
1490                    file.url_path,
1491                    axum::routing::get(move || async move {
1492                        use axum::response::IntoResponse;
1493                        let cc = file.cache_control.unwrap_or("public, max-age=86400");
1494                        axum::http::Response::builder()
1495                            .status(axum::http::StatusCode::OK)
1496                            .header(axum::http::header::CONTENT_TYPE, file.content_type)
1497                            .header(axum::http::header::CACHE_CONTROL, cc)
1498                            .body(axum::body::Body::from(file.body))
1499                            .unwrap_or_else(|_| {
1500                                axum::http::StatusCode::INTERNAL_SERVER_ERROR.into_response()
1501                            })
1502                    }),
1503                );
1504            }
1505        }
1506
1507        // Phase 5.45 — mount the unified static pipeline handler. Walk
1508        // every plugin's `static_dirs()` into a namespace -> source_dir
1509        // registry (a duplicate namespace fails the build loudly), then
1510        // nest ONE handler at the configured `static_url` base. It
1511        // resolves `/static/<ns>/<rest>` live-from-source in dev and
1512        // from `static_root` in prod (see `crate::static_files`).
1513        //
1514        // This coexists with the `StaticFile` embedded routes mounted in
1515        // Phase 5.4 above — embedded assets stay the zero-config default;
1516        // the filesystem handler is additive.
1517        //
1518        // A CDN-style `static_url` (an absolute http(s):// origin) can't
1519        // be nested as a local route prefix; in that mode assets are
1520        // served off the CDN and the local handler is intentionally not
1521        // mounted — the `static()` template helper still emits the
1522        // absolute URLs.
1523        let settings = crate::settings::get();
1524        let static_base = settings.static_url.trim_end_matches('/');
1525        let is_cdn_url = settings.static_url.starts_with("http://")
1526            || settings.static_url.starts_with("https://")
1527            || settings.static_url.starts_with("//");
1528
1529        // App/site-level static dirs served at the bare `static_url` root.
1530        // A `StoragePlugin`'s static side mounted AT `static_url` contributes its
1531        // directory here (and skips nesting its own catch-all), so the
1532        // framework owns `static_url` as ONE mount — a second
1533        // `/static/{*rest}` nest is exactly the conflict this avoids.
1534        let root_dirs = crate::static_files::StaticContribution::collect_root_dirs(&sorted_plugins);
1535
1536        // Publish the static contributions ambiently for `collectstatic`
1537        // (the `StoragePlugin` CLI command). Published UNCONDITIONALLY —
1538        // before the serving-mode gate below — because `collectstatic`
1539        // copies assets to disk regardless of serving mode (a CDN-mode
1540        // app still needs the disk tree built for upload). Mirrors the
1541        // `settings` ambient OnceLock: read-only config set once at build.
1542        crate::static_files::publish_static(crate::static_files::PublishedStatic {
1543            contributions: crate::static_files::StaticContribution::collect(&sorted_plugins),
1544            root_dirs: root_dirs.clone(),
1545        });
1546
1547        // Load the hashed-asset manifest (`<static_root>/staticfiles.json`)
1548        // if `collectstatic --hashed` has produced one. With a manifest
1549        // present, `resolve_static_url` / the `static()` template global
1550        // emit content-hashed URLs so prod assets carry far-future cache
1551        // headers. Absent (no `--hashed` run), this is a no-op and URLs
1552        // stay plain. Loaded unconditionally — the URL resolution applies
1553        // whether or not this app serves the bytes itself.
1554        crate::static_files::load_manifest(&settings.static_root);
1555
1556        if !is_cdn_url && !static_base.is_empty() {
1557            let registry = crate::static_files::StaticRegistry::from_plugins(&sorted_plugins)
1558                .map_err(|c| BuildError::DuplicateStaticNamespace {
1559                    namespace: c.namespace,
1560                    first_plugin: c.first_plugin,
1561                    second_plugin: c.second_plugin,
1562                })?;
1563            // Nothing to serve and no app static dirs — don't claim the
1564            // `static_url` path at all, so a consumer that wants to mount
1565            // their own router there can.
1566            if !registry.is_empty() || !root_dirs.is_empty() {
1567                let state = crate::static_files::StaticHandlerState {
1568                    registry,
1569                    static_root: std::path::PathBuf::from(&settings.static_root),
1570                    root_dirs,
1571                    dev: matches!(settings.environment, crate::settings::Environment::Dev),
1572                };
1573                let static_router = Router::new()
1574                    .fallback(crate::static_files::static_handler)
1575                    .with_state(state);
1576                router = router.nest_service(static_base, static_router);
1577            }
1578        }
1579
1580        // Phase 5.5 — apply each plugin's middleware in topological
1581        // order. Later plugins wrap earlier ones, so a security
1582        // plugin declared after the auth plugin sees the auth-
1583        // augmented router and can add its own layer on top. This
1584        // is the M7 deferral being lifted now that umbral-security
1585        // needs it.
1586        for plugin in &sorted_plugins {
1587            router = plugin.wrap_router(router);
1588        }
1589
1590        // Phase 5.6 — install the 404 fallback. Four cases:
1591        //
1592        // 1. slash_redirect = Off, not_found_template = None, default pages off:
1593        //    no-op. axum's built-in empty 404 is what users see.
1594        // 2. slash_redirect = Off, not_found_template = None, default pages ON:
1595        //    install the not-found fallback; render_not_found will use the
1596        //    embedded default_404 template.
1597        // 3. slash_redirect = Off, not_found_template = Some(name):
1598        //    install the not-found fallback directly. Renders the
1599        //    template on every miss.
1600        // 4. slash_redirect != Off:
1601        //    install the slash-redirect fallback. It handles its own
1602        //    404 path internally — when no alternate matches, it
1603        //    renders the configured not-found template (or the default
1604        //    if enabled, or plain text if both are absent).
1605        //
1606        // The slash-redirect fallback ALWAYS captures a router
1607        // snapshot taken BEFORE the fallback is installed, so the
1608        // alternate-path probe can't recursively re-hit the fallback.
1609        let need_not_found_fallback = self.not_found_template.is_some() || self.default_error_pages;
1610        match (self.slash_redirect, need_not_found_fallback) {
1611            (crate::slash::SlashRedirect::Off, false) => {
1612                // axum's default 404 — nothing to do.
1613            }
1614            (crate::slash::SlashRedirect::Off, true) => {
1615                let fallback = crate::errors::not_found_fallback(self.not_found_template.clone());
1616                router = router.fallback(fallback);
1617            }
1618            (policy, _) => {
1619                let snapshot = router.clone();
1620                let fallback = crate::slash::slash_redirect_fallback(
1621                    snapshot,
1622                    policy,
1623                    self.not_found_template.clone(),
1624                );
1625                router = router.fallback(fallback);
1626            }
1627        }
1628
1629        // Phase 5.65 — framework middleware stack (feature #68). App-level
1630        // middleware first, then every plugin's contribution in topological
1631        // order, collected into one stack and installed as a single layer.
1632        // Placed AFTER the 404 fallback so middleware sees misses too, and
1633        // BEFORE the panic / compression / CORS / host layers so those stay
1634        // the outermost wrappers (security and content-encoding run before
1635        // user middleware ever touches the request).
1636        let mut middleware_stack = crate::middleware::MiddlewareStack::new();
1637        middleware_stack.extend(std::mem::take(&mut self.middleware));
1638        for plugin in &sorted_plugins {
1639            middleware_stack.extend(plugin.middleware());
1640        }
1641        router = middleware_stack.apply(router);
1642
1643        // Phase 5.66 — request-scoped routing context (DatabaseRouter
1644        // foundation). When a resolver is registered, wrap the whole
1645        // downstream future in `route_context::scope`. Installed OUTSIDE the
1646        // middleware stack above so the task-local is established before any
1647        // middleware or handler runs — every `.await` in the request,
1648        // including ORM calls that read `route_context::current()`, then sees
1649        // the resolved context. A `from_fn` layer is the only mechanism that
1650        // can wrap `next.run(req)` in a scope; the `Middleware` contract's
1651        // `before_request(req) -> req` cannot.
1652        if let Some(resolver) = self.route_context_resolver.take() {
1653            router = router.layer(axum::middleware::from_fn_with_state(
1654                resolver,
1655                route_context_scope_layer,
1656            ));
1657        }
1658
1659        // Phase 5.7 — wrap with the panic-catch layer. Comes AFTER the
1660        // fallback wiring so a panicking fallback handler is also caught
1661        // (the panic-catch layer wraps the entire router).
1662        //
1663        // Always installed when: a user-supplied server_error_template is
1664        // set, OR default pages are enabled (the embedded default_500 fires
1665        // in that case), OR an on_server_error hook is registered.
1666        let need_panic_layer = self.server_error_template.is_some()
1667            || self.default_error_pages
1668            || self.server_error_hook.is_some();
1669        if need_panic_layer {
1670            let handler = crate::errors::server_error_panic_handler(
1671                self.server_error_template.clone(),
1672                self.server_error_hook.clone(),
1673            );
1674            router = router.layer(tower_http::catch_panic::CatchPanicLayer::custom(handler));
1675
1676            // Phase 5.8 — wrap with the response-rendering middleware so
1677            // any 500 produced by a handler (not just a panic) gets
1678            // re-rendered through the configured 500 template. The
1679            // middleware checks Content-Type: HTML responses (from the
1680            // panic handler above, or from a handler that rendered its
1681            // own template) pass through; plain-text 500s get re-rendered.
1682            // Also fires `on_server_error` for handler-Err paths.
1683            let render_state = crate::errors::Render500State {
1684                template: self.server_error_template.clone(),
1685                hook: self.server_error_hook.clone(),
1686            };
1687            router = router.layer(axum::middleware::from_fn_with_state(
1688                render_state,
1689                crate::errors::render_500_middleware,
1690            ));
1691        }
1692
1693        // General custom error pages: style any registered status code
1694        // (429/403/410/…) the way the 500 path does, for handler-Err
1695        // responses — rendering each through its template while preserving the
1696        // status. Already-HTML and unregistered statuses pass through; this is
1697        // independent of the 500 layer above (different status codes).
1698        if !self.error_templates.is_empty() {
1699            let state = crate::errors::RenderErrorState {
1700                templates: std::sync::Arc::new(std::mem::take(&mut self.error_templates)),
1701            };
1702            router = router.layer(axum::middleware::from_fn_with_state(
1703                state,
1704                crate::errors::render_error_middleware,
1705            ));
1706        }
1707
1708        // Optional response compression (gzip / brotli), opt-in via
1709        // `AppBuilder::compression`. tower-http chooses the algorithm from
1710        // `Accept-Encoding` and skips already-encoded / non-compressible
1711        // bodies. Applied here so it wraps handler responses; CORS + host
1712        // checks layer outside it.
1713        if self.compress {
1714            router = router.layer(tower_http::compression::CompressionLayer::new());
1715        }
1716
1717        // Phase 5.9 — CORS, applied last so it's the outermost
1718        // wrapper. Preflight `OPTIONS` is answered before any
1719        // plugin/handler sees the request; response headers are
1720        // added on the way back out regardless of which downstream
1721        // layer produced the body.
1722        if let Some(cors) = self.cors.take() {
1723            router = router.layer(cors.into_layer());
1724        }
1725        // Path-scoped CORS (e.g. `/api`) — layered after the global one so each
1726        // only touches responses for requests under its prefix.
1727        for (prefix, config) in std::mem::take(&mut self.cors_scoped) {
1728            router = router.layer(crate::cors::ScopedCorsLayer::new(
1729                prefix,
1730                config.into_layer(),
1731            ));
1732        }
1733
1734        // Phase 5.95 — Host-header validation (allowed-hosts allowlist). Applied
1735        // outermost so a forged `Host` is rejected with a 400 before any
1736        // handler, plugin, or CORS logic runs. Enforced only in
1737        // `Environment::Prod`; dev passes through. Allowlist is
1738        // `settings.allowed_hosts` (`"*"` disables; `.example.com` = subdomain).
1739        let host_policy = crate::hosts::HostPolicy::new(
1740            &settings.allowed_hosts,
1741            matches!(settings.environment, crate::settings::Environment::Prod),
1742        );
1743        router = router.layer(axum::middleware::from_fn_with_state(
1744            host_policy,
1745            crate::hosts::host_guard,
1746        ));
1747
1748        // Request hardening (audit_2 core-web H11) — a framework-wide body-size
1749        // cap and a per-request timeout, both safe-by-default and opt-out-able
1750        // via `AppBuilder::max_request_body` / `request_timeout`. Layered
1751        // outermost (just under the trace span) so they bound EVERY request —
1752        // including host-guard rejections — before an inner extractor or the
1753        // multipart parser can buffer an oversized body or a hung handler can
1754        // pin a task. `RequestBodyLimitLayer` returns 413; `TimeoutLayer`
1755        // returns 408.
1756        if let Some(limit) = self.max_request_body_bytes {
1757            router = router.layer(tower_http::limit::RequestBodyLimitLayer::new(limit));
1758        }
1759        if let Some(timeout) = self.request_timeout {
1760            router = router.layer(tower_http::timeout::TimeoutLayer::with_status_code(
1761                axum::http::StatusCode::REQUEST_TIMEOUT,
1762                timeout,
1763            ));
1764        }
1765
1766        // audit_2 H10 — minimal hardening response headers from core, so a
1767        // default app that forgot SecurityPlugin is still not clickjackable /
1768        // MIME-sniffable. Set ONLY if absent (SecurityPlugin's configured
1769        // values win), and applied outer to the host/limit/timeout layers so
1770        // their 4xx responses carry them too. HSTS is deliberately NOT set here
1771        // — it's sticky and subdomain-scoped, so it stays SecurityPlugin's
1772        // configurable responsibility.
1773        if self.default_security_headers {
1774            router = router.layer(axum::middleware::from_fn(default_security_headers_layer));
1775        }
1776
1777        // Phase 5.99 — request tracing span. Applied outermost so every request
1778        // (including host-guard rejections) runs inside a span. The span
1779        // carries `http.method`, `http.route`/`uri`, and the response
1780        // `http.status_code`; this is what an OpenTelemetry layer (installed by
1781        // an app via `umbral_logs::observability::init`) exports as one span per
1782        // request. Without an OTel layer attached it's a cheap `tracing` span
1783        // that the fmt subscriber can surface under `RUST_LOG=tower_http=debug`.
1784        // W3C `traceparent` propagation (extracting an upstream trace context
1785        // from the inbound header) is a noted follow-up; this layer creates the
1786        // local request span.
1787        router = router.layer(
1788            tower_http::trace::TraceLayer::new_for_http().make_span_with(
1789                |request: &axum::http::Request<axum::body::Body>| {
1790                    tracing::info_span!(
1791                        "http.request",
1792                        http.method = %request.method(),
1793                        http.route = %request.uri().path(),
1794                        http.status_code = tracing::field::Empty,
1795                    )
1796                },
1797            ),
1798        );
1799
1800        // Phase 6 — `on_ready` USED to fire here. It doesn't any more: the hooks
1801        // seed content and backfill rows, and `build()` runs before the CLI has
1802        // parsed argv, so `migrate` against a fresh database ran every seed
1803        // before a single table existed (gaps3 #41). The caller now decides when
1804        // the app is ready; see [`App::ready`], which `serve()` and
1805        // `umbral_cli::dispatch` call at the right moment.
1806        Ok(App {
1807            router,
1808            plugins: sorted_plugins,
1809            commands: self.commands,
1810            auto_migrate_on_serve: self.auto_migrate_on_serve,
1811            drain_delay: self.drain_delay,
1812            ready_fired: std::sync::atomic::AtomicBool::new(false),
1813        })
1814    }
1815}
1816
1817/// The axum middleware fn installed by [`AppBuilder::route_context`]: run the
1818/// resolver against the incoming request to build a [`crate::db::RouteContext`],
1819/// then drive the ENTIRE downstream future inside
1820/// [`crate::db::route_context::scope`]. Scoping `next.run(req)` (rather than
1821/// just a prefix of it) is what keeps the task-local alive across every
1822/// `.await` the handler performs, so ambient ORM calls route per the resolved
1823/// context.
1824async fn route_context_scope_layer(
1825    axum::extract::State(resolver): axum::extract::State<RouteContextResolver>,
1826    req: crate::web::Request,
1827    next: axum::middleware::Next,
1828) -> crate::web::Response {
1829    let ctx = resolver(&req);
1830    crate::db::route_context::scope(ctx, next.run(req)).await
1831}
1832
1833/// Resolve when the process receives a shutdown signal — `SIGTERM` (the deploy
1834/// / container-stop signal) or `SIGINT` (Ctrl-C). Drives `serve`'s graceful
1835/// shutdown (audit_2 core-app-config #13). On non-Unix only Ctrl-C is wired.
1836async fn shutdown_signal() {
1837    let ctrl_c = async {
1838        let _ = tokio::signal::ctrl_c().await;
1839    };
1840
1841    #[cfg(unix)]
1842    let terminate = async {
1843        match tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate()) {
1844            Ok(mut sig) => {
1845                sig.recv().await;
1846            }
1847            // If the handler can't be installed, never fire this arm.
1848            Err(_) => std::future::pending::<()>().await,
1849        }
1850    };
1851    #[cfg(not(unix))]
1852    let terminate = std::future::pending::<()>();
1853
1854    tokio::select! {
1855        _ = ctrl_c => {}
1856        _ = terminate => {}
1857    }
1858    tracing::info!("umbral: shutdown signal received; finishing in-flight requests");
1859}
1860
1861/// The future `serve` hands to axum's `with_graceful_shutdown` (Kikosi #5).
1862///
1863/// Awaits `signal`, marks the process draining so readiness probes go 503, then
1864/// — if `drain_delay` is non-zero — keeps serving for that long before resolving,
1865/// which is when axum stops accepting new connections. The delay is the window a
1866/// load balancer uses to observe the 503 and route new traffic elsewhere.
1867///
1868/// Generic over the signal future so the drain sequencing is testable without
1869/// actually delivering a `SIGTERM`.
1870async fn drain_after<F: std::future::Future>(signal: F, drain_delay: std::time::Duration) {
1871    signal.await;
1872    crate::shutdown::begin_drain();
1873    if !drain_delay.is_zero() {
1874        tracing::info!(
1875            drain_secs = drain_delay.as_secs_f64(),
1876            "umbral: draining — readiness now reports 503; holding before stop"
1877        );
1878        tokio::time::sleep(drain_delay).await;
1879    }
1880}
1881
1882/// audit_2 H19 — warn about the app's own mutating routes that carry no
1883/// recorded permission. A default-DENY router is a future-major change; this
1884/// boot Warning is the non-breaking default: it makes a forgotten
1885/// authorization gate visible at boot instead of shipping as an open endpoint.
1886/// [`AppBuilder::deny_ungated_mutations`] promotes the same finding to a hard
1887/// [`BuildError::UngatedMutatingRoutes`] for apps that want it enforced.
1888///
1889/// Scope + honesty: only routes registered through `Routes` (the app's
1890/// `.routes(...)`) are checked — plugin routes gate via their own conventions.
1891/// A route gated by a hand-applied `.layer(permission_required(...))` is opaque
1892/// to `RouteSpec`, so it can't be distinguished from an ungated one; the
1893/// warning says so and points at the `require_permission(...)` builder (which
1894/// records the permission). An intentionally-public route is a false positive
1895/// the operator ignores.
1896fn warn_ungated_mutating_routes(ungated: &[String]) {
1897    tracing::warn!(
1898        "audit_2 H19: {} app mutating route(s) have no recorded permission: [{}]. \
1899         Gate them with the umbral-permissions `Routes::require_permission(...)` builder \
1900         so the framework records the permission (a hand-applied \
1901         `.layer(permission_required(...))` is NOT visible to this audit — prefer the \
1902         builder). If a route is intentionally public, ignore this. To make this a hard \
1903         boot error instead, call `App::builder().deny_ungated_mutations()`.",
1904        ungated.len(),
1905        ungated.join(", ")
1906    );
1907}
1908
1909/// The pure core of the H19 audit: the `"METHOD /path"`
1910/// labels of every route with a mutating method and no recorded permission.
1911/// Split out so the audit's selection logic is unit-testable without a live
1912/// `App::build()` / tracing subscriber.
1913fn ungated_mutating_routes(specs: &[crate::routes::RouteSpec]) -> Vec<String> {
1914    const MUTATING: [&str; 4] = ["POST", "PUT", "PATCH", "DELETE"];
1915    specs
1916        .iter()
1917        .filter(|s| s.permission.is_none() && s.methods.iter().any(|m| MUTATING.contains(m)))
1918        .map(|s| format!("{} {}", s.methods.join("/"), s.path))
1919        .collect()
1920}
1921
1922/// Set minimal hardening response headers, each ONLY if the response doesn't
1923/// already carry it — so `SecurityPlugin` (or a handler) can override, and no
1924/// header is ever duplicated (audit_2 H10).
1925async fn default_security_headers_layer(
1926    req: crate::web::Request,
1927    next: axum::middleware::Next,
1928) -> crate::web::Response {
1929    use axum::http::HeaderValue;
1930    use axum::http::header::{
1931        HeaderName, REFERRER_POLICY, X_CONTENT_TYPE_OPTIONS, X_FRAME_OPTIONS,
1932    };
1933
1934    let mut resp = next.run(req).await;
1935    let headers = resp.headers_mut();
1936    let mut set_if_absent = |name: HeaderName, value: &'static str| {
1937        if !headers.contains_key(&name) {
1938            headers.insert(name, HeaderValue::from_static(value));
1939        }
1940    };
1941    set_if_absent(X_CONTENT_TYPE_OPTIONS, "nosniff");
1942    set_if_absent(X_FRAME_OPTIONS, "DENY");
1943    set_if_absent(REFERRER_POLICY, "strict-origin-when-cross-origin");
1944    resp
1945}
1946
1947/// One cross-plugin foreign-key edge derived from the model registry.
1948///
1949/// `plugin`'s table `table` carries a physical `REFERENCES "<fk_target>"`, and
1950/// `fk_target` is owned by `depends_on`. Reported on
1951/// [`BuildError::ForeignKeyCycle`] so the operator sees the *column* that
1952/// forced the ordering, not just two plugin names.
1953#[derive(Debug, Clone, PartialEq, Eq)]
1954pub struct FkEdge {
1955    /// The plugin whose table holds the foreign key.
1956    pub plugin: &'static str,
1957    /// The plugin that owns the referenced table, and so must migrate first.
1958    pub depends_on: &'static str,
1959    /// The referencing table (`Model::TABLE` of the model holding the FK).
1960    pub table: String,
1961    /// The referenced table (`fk_target` of the FK column).
1962    pub fk_target: String,
1963}
1964
1965/// Derive the cross-plugin ordering edges the *schema already states*.
1966///
1967/// `Plugin::dependencies()` is the edge set an author declares by hand. This is
1968/// the edge set the models spell out: a `ForeignKey<T>` field renders
1969/// `REFERENCES "<T::TABLE>"` inside `CREATE TABLE`, so the plugin owning
1970/// `T::TABLE` must create it first. Postgres enforces the target's existence at
1971/// `CREATE TABLE` time; SQLite silently accepts the dangling reference, which is
1972/// why omitting the edge only ever failed on a fresh Postgres (gaps3 #40).
1973///
1974/// Two exclusions, both deliberate:
1975///
1976/// - **Same-plugin FKs** impose no *plugin* ordering. Column order inside one
1977///   plugin's migration is the diff engine's problem, not the sort's.
1978/// - **`#[umbral(db_constraint = false)]`** renders no `REFERENCES` clause (the
1979///   only valid shape for a cross-database FK, gaps2 #22), so it creates no DDL
1980///   ordering obligation.
1981///
1982/// An FK whose target table belongs to no registered plugin — an app-owned model
1983/// from `.model::<T>()` — yields no edge: the implicit `"app"` plugin is pinned
1984/// last by `App::build()` precisely because app models FK *into* plugin tables.
1985fn fk_plugin_edges(plugins: &[Box<dyn Plugin>]) -> Vec<FkEdge> {
1986    use std::collections::BTreeMap;
1987
1988    // table -> owning plugin. Built across every registered plugin first so a
1989    // forward reference (a plugin FK-ing a table owned by a plugin registered
1990    // later) still resolves.
1991    let mut owner_of_table: BTreeMap<String, &'static str> = BTreeMap::new();
1992    for plugin in plugins {
1993        for model in plugin.models() {
1994            owner_of_table.insert(model.table, plugin.name());
1995        }
1996    }
1997
1998    let mut edges: Vec<FkEdge> = Vec::new();
1999    for plugin in plugins {
2000        let name = plugin.name();
2001        for model in plugin.models() {
2002            for column in &model.fields {
2003                let Some(target) = column.fk_target.as_deref() else {
2004                    continue;
2005                };
2006                if !column.db_constraint {
2007                    continue;
2008                }
2009                let Some(&owner) = owner_of_table.get(target) else {
2010                    continue;
2011                };
2012                if owner == name {
2013                    continue;
2014                }
2015                edges.push(FkEdge {
2016                    plugin: name,
2017                    depends_on: owner,
2018                    table: model.table.clone(),
2019                    fk_target: target.to_string(),
2020                });
2021            }
2022        }
2023    }
2024    edges
2025}
2026
2027/// Kahn's algorithm over `deps` (plugin -> the set it waits on), with a
2028/// name-sorted ready queue so ties resolve deterministically. Returns the
2029/// topological order, or the still-unsorted names when the graph has a cycle.
2030fn toposort(
2031    mut remaining_deps: std::collections::BTreeMap<
2032        &'static str,
2033        std::collections::BTreeSet<&'static str>,
2034    >,
2035) -> Result<Vec<&'static str>, Vec<&'static str>> {
2036    use std::collections::BTreeSet;
2037
2038    let mut ready: BTreeSet<&'static str> = remaining_deps
2039        .iter()
2040        .filter_map(|(name, deps)| if deps.is_empty() { Some(*name) } else { None })
2041        .collect();
2042
2043    let mut order: Vec<&'static str> = Vec::with_capacity(remaining_deps.len());
2044    while let Some(name) = ready.iter().next().copied() {
2045        ready.remove(&name);
2046        remaining_deps.remove(&name);
2047        order.push(name);
2048        for (other_name, deps) in remaining_deps.iter_mut() {
2049            if deps.remove(&name) && deps.is_empty() {
2050                ready.insert(*other_name);
2051            }
2052        }
2053    }
2054
2055    if remaining_deps.is_empty() {
2056        Ok(order)
2057    } else {
2058        Err(remaining_deps.keys().copied().collect())
2059    }
2060}
2061
2062/// Validate the registered plugins and return them in a stable
2063/// topological order. Standard Kahn's algorithm with a name-sorted ready queue
2064/// so ties resolve deterministically.
2065///
2066/// The edge set is the union of two sources:
2067///
2068/// 1. `Plugin::dependencies()` — what the author declared.
2069/// 2. Cross-plugin foreign keys read off the models ([`fk_plugin_edges`]) — what
2070///    the schema already states.
2071///
2072/// Before gaps3 #40 only (1) fed the sort, so an app where *no* plugin declared
2073/// a dependency had every plugin at in-degree 0 and the "topological" order
2074/// collapsed to alphabetical. `"accounts"` sorts before `"auth"`, and its
2075/// `CREATE TABLE ... REFERENCES "auth_user"` ran against a database with no
2076/// `auth_user`. Declaring your own dependencies is still the plugin author's job;
2077/// the framework just no longer lets the omission reach production silently.
2078///
2079/// Rejects:
2080///
2081/// - A plugin claiming the reserved `"app"` name.
2082/// - Two plugins reporting the same `name()`.
2083/// - A `dependencies()` entry that doesn't name a registered plugin.
2084/// - A declared dependency cycle (`BuildError::PluginCycle`).
2085/// - A cycle introduced by the foreign keys themselves
2086///   (`BuildError::ForeignKeyCycle`, which names the offending columns).
2087fn sort_plugins(plugins: Vec<Box<dyn Plugin>>) -> Result<Vec<Box<dyn Plugin>>, BuildError> {
2088    use std::collections::{BTreeMap, BTreeSet};
2089
2090    // Reserved + duplicate-name checks. The implicit `"app"` plugin is
2091    // not counted toward duplicates; only the user's plugin list is.
2092    let mut seen: BTreeSet<&'static str> = BTreeSet::new();
2093    for plugin in &plugins {
2094        let name = plugin.name();
2095        if name == crate::migrate::APP_PLUGIN_NAME {
2096            return Err(BuildError::ReservedPluginName);
2097        }
2098        if !seen.insert(name) {
2099            return Err(BuildError::DuplicatePluginName { name });
2100        }
2101    }
2102
2103    // Index plugins by name for the dependency lookups + the
2104    // sort-by-name traversal below. We pull the boxes out of the
2105    // input vec by index later, so the index table stays alongside.
2106    let by_name: BTreeMap<&'static str, usize> = plugins
2107        .iter()
2108        .enumerate()
2109        .map(|(i, p)| (p.name(), i))
2110        .collect();
2111
2112    // Dependency-exists check. Done before the toposort so a missing
2113    // dep surfaces with the asking plugin's name attached, not as a
2114    // cycle false-positive.
2115    for plugin in &plugins {
2116        for dep in plugin.dependencies() {
2117            if !by_name.contains_key(dep) {
2118                return Err(BuildError::DependencyNotFound {
2119                    plugin: plugin.name(),
2120                    missing: dep,
2121                });
2122            }
2123        }
2124    }
2125
2126    let declared: BTreeMap<&'static str, BTreeSet<&'static str>> = plugins
2127        .iter()
2128        .map(|p| (p.name(), p.dependencies().iter().copied().collect()))
2129        .collect();
2130
2131    let fk_edges = fk_plugin_edges(&plugins);
2132    let mut combined = declared.clone();
2133    for edge in &fk_edges {
2134        combined
2135            .get_mut(edge.plugin)
2136            .expect("every FK edge names a registered plugin")
2137            .insert(edge.depends_on);
2138    }
2139
2140    let order = match toposort(combined) {
2141        Ok(order) => order,
2142        Err(stuck) => {
2143            // The combined graph cycles. Re-run on the declared edges alone to
2144            // find out who is to blame. If the declared graph is acyclic, the
2145            // foreign keys introduced the cycle — report the columns that did
2146            // it rather than a bare `PluginCycle` the author never wrote.
2147            //
2148            // Across crates this is unreachable: `ForeignKey<T>` needs `T` in
2149            // scope, so mutually-referencing plugin crates would be a circular
2150            // Cargo dependency. Two plugins defined in ONE crate can still do
2151            // it, and a cross-plugin FK cycle has no valid `CREATE TABLE` order
2152            // on a fresh database either way.
2153            if toposort(declared).is_ok() {
2154                let stuck: BTreeSet<&'static str> = stuck.into_iter().collect();
2155                let edges: Vec<FkEdge> = fk_edges
2156                    .into_iter()
2157                    .filter(|e| stuck.contains(e.plugin) && stuck.contains(e.depends_on))
2158                    .collect();
2159                return Err(BuildError::ForeignKeyCycle { edges });
2160            }
2161            return Err(BuildError::PluginCycle { names: stuck });
2162        }
2163    };
2164
2165    // Reorder the owned boxes into topological order. We pull each
2166    // plugin out of an `Option` slot so the move is statically
2167    // tracked; every slot is taken exactly once because the toposort
2168    // produced one entry per plugin.
2169    let mut slots: Vec<Option<Box<dyn Plugin>>> = plugins.into_iter().map(Some).collect();
2170    let mut sorted: Vec<Box<dyn Plugin>> = Vec::with_capacity(order.len());
2171    for name in order {
2172        let idx = by_name[&name];
2173        sorted.push(
2174            slots[idx]
2175                .take()
2176                .expect("toposort produced one entry per plugin"),
2177        );
2178    }
2179    Ok(sorted)
2180}
2181
2182/// Errors that can occur during `AppBuilder::build()`.
2183#[derive(Debug)]
2184pub enum BuildError {
2185    /// `.settings(Settings)` wasn't called on the builder.
2186    SettingsMissing,
2187    /// `.database("default", pool)` wasn't called on the builder.
2188    DefaultPoolMissing,
2189    /// The URL scheme in `settings.database_url` doesn't match any
2190    /// shipped backend.
2191    BackendDetect(crate::backend::BackendDetectError),
2192    /// One or more system checks failed with `Severity::Error`. The
2193    /// full list of findings is in the variant.
2194    SystemCheckFailed {
2195        findings: Vec<crate::check::SystemCheckFinding>,
2196    },
2197    /// A plugin's `dependencies()` lists a plugin that was never
2198    /// registered with `.plugin(...)`. Carries the unmet name plus
2199    /// the plugin that asked for it.
2200    DependencyNotFound {
2201        plugin: &'static str,
2202        missing: &'static str,
2203    },
2204    /// The dependency graph has a cycle. Carries the plugin names that
2205    /// form it (in any cyclic order; the diagnostic is "these N plugins
2206    /// reference each other").
2207    PluginCycle { names: Vec<&'static str> },
2208    /// The plugins' *foreign keys* form a cycle, so no `CREATE TABLE` order
2209    /// satisfies every `REFERENCES` clause on a fresh database. Distinct from
2210    /// [`BuildError::PluginCycle`], which reports a cycle the author declared
2211    /// via `dependencies()`; here nothing was declared and the cycle is implied
2212    /// by the models. Carries the FK edges that close the loop so the message
2213    /// can name the columns, not just the plugins.
2214    ForeignKeyCycle { edges: Vec<FkEdge> },
2215    /// Two registered plugins share a `name()`. Plugin names are keys
2216    /// in the migration tracking table and the dependency graph; a
2217    /// collision would break both.
2218    DuplicatePluginName { name: &'static str },
2219    /// A plugin claimed the reserved `"app"` name (used by the
2220    /// implicit plugin that owns `.model::<T>()` registrations).
2221    ReservedPluginName,
2222    /// A plugin's `on_ready` returned an error. Carries the plugin's
2223    /// name plus the underlying error.
2224    PluginOnReady {
2225        plugin: &'static str,
2226        source: Box<dyn std::error::Error + Send + Sync>,
2227    },
2228    /// The templates engine failed to initialise. Carries the
2229    /// underlying `TemplateError` (an IO error reading a template
2230    /// file, or a syntax error in one of the loaded templates).
2231    TemplatesInit(crate::templates::TemplateError),
2232    /// A plugin's `database()` returned an alias that isn't in the
2233    /// registered pool set. Surfaces a typo at boot with a clear
2234    /// "register the pool first" diagnostic instead of letting
2235    /// `db::pool_for` panic at first query.
2236    PluginDatabaseAlias {
2237        plugin: &'static str,
2238        alias: &'static str,
2239    },
2240    /// A `settings.databases` entry could not be opened as a lazy pool at boot
2241    /// (audit_2 H17) — e.g. an unsupported URL scheme. Carries the alias and the
2242    /// sqlx error.
2243    SettingsDatabasePool { alias: String, error: sqlx::Error },
2244    /// The URL-derived backend (from `settings.database_url`) doesn't
2245    /// match the runtime type of the default pool passed to
2246    /// `.database("default", ...)`. Catches the case where the URL
2247    /// says `postgres://` but a `SqlitePool` was registered, or vice
2248    /// versa.
2249    DatabaseBackendMismatch {
2250        url_backend: &'static str,
2251        pool_backend: &'static str,
2252    },
2253    /// A foreign key targets a model on a different database than the
2254    /// model that declares it, and the field has NOT opted out of the
2255    /// physical constraint. A `REFERENCES` clause can't span databases,
2256    /// so this would emit invalid DDL. Fix by either routing both
2257    /// models to the same database, or marking the FK
2258    /// `#[umbral(db_constraint = false)]` to keep it a logical-only
2259    /// relation. Closes gaps2 #22.
2260    CrossDatabaseForeignKey {
2261        model: &'static str,
2262        field: &'static str,
2263        model_db: &'static str,
2264        target_db: &'static str,
2265    },
2266    /// Two plugins declared the same static namespace via
2267    /// `Plugin::static_dirs()`. Namespaces are the per-plugin URL/disk
2268    /// segment under `static_url` / `static_root`; a collision would
2269    /// silently shadow one plugin's assets with another's, so the build
2270    /// fails loudly and names both plugins.
2271    DuplicateStaticNamespace {
2272        namespace: &'static str,
2273        first_plugin: &'static str,
2274        second_plugin: &'static str,
2275    },
2276    /// `.deny_ungated_mutations()` was set and one or more app-level mutating
2277    /// routes (POST/PUT/PATCH/DELETE registered via `.routes(...)`) carry no
2278    /// recorded permission (gaps3 #28 P1, enforcing the audit_2 H19 audit).
2279    /// Carries the `"METHOD /path"` label of each offending route. Fix by gating
2280    /// them with the umbral-permissions `Routes::require_permission(...)` builder
2281    /// (which records the permission), or drop the strict flag if a route is
2282    /// intentionally public.
2283    UngatedMutatingRoutes { routes: Vec<String> },
2284}
2285
2286impl std::fmt::Display for BuildError {
2287    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
2288        match self {
2289            BuildError::SettingsMissing => write!(
2290                f,
2291                "umbral: App::builder() requires Settings; call .settings(Settings::from_env()?) before .build()"
2292            ),
2293            BuildError::BackendDetect(err) => write!(f, "{err}"),
2294            BuildError::SystemCheckFailed { findings } => {
2295                writeln!(f, "umbral: {} system check(s) failed:", findings.len())?;
2296                for finding in findings {
2297                    write!(f, "  - [{}] {}", finding.check_id, finding.message)?;
2298                    if let Some(hint) = &finding.hint {
2299                        write!(f, " (hint: {hint})")?;
2300                    }
2301                    writeln!(f)?;
2302                }
2303                Ok(())
2304            }
2305            BuildError::DefaultPoolMissing => write!(
2306                f,
2307                "umbral: App::builder() requires a default DB pool; call .database(\"default\", umbral::db::connect(&url).await?) before .build()"
2308            ),
2309            BuildError::DependencyNotFound { plugin, missing } => write!(
2310                f,
2311                "umbral: plugin `{plugin}` depends on `{missing}`, which isn't registered; \
2312                 call .plugin({missing}::default()) on the builder"
2313            ),
2314            BuildError::PluginCycle { names } => {
2315                write!(f, "umbral: plugin dependency cycle: {}", names.join(" -> "))
2316            }
2317            BuildError::ForeignKeyCycle { edges } => {
2318                writeln!(
2319                    f,
2320                    "umbral: the plugins' foreign keys form a cycle, so no CREATE TABLE order \
2321                     satisfies every REFERENCES clause on a fresh database:"
2322                )?;
2323                for edge in edges {
2324                    writeln!(
2325                        f,
2326                        "  `{}`.\"{}\" REFERENCES \"{}\", owned by `{}`",
2327                        edge.plugin, edge.table, edge.fk_target, edge.depends_on
2328                    )?;
2329                }
2330                write!(
2331                    f,
2332                    "break the cycle by making one side a nullable FK added in a later \
2333                     migration, or by opting that column out of the physical constraint with \
2334                     #[umbral(db_constraint = false)]"
2335                )
2336            }
2337            BuildError::DuplicatePluginName { name } => write!(
2338                f,
2339                "umbral: two plugins both report name `{name}`; plugin names are unique keys \
2340                 (migration tracking, dependency graph)"
2341            ),
2342            BuildError::SettingsDatabasePool { alias, error } => write!(
2343                f,
2344                "umbral: could not open the `settings.databases` pool for alias `{alias}`: \
2345                 {error}"
2346            ),
2347            BuildError::ReservedPluginName => write!(
2348                f,
2349                "umbral: the plugin name `app` is reserved for models registered via \
2350                 .model::<T>(); pick a different name"
2351            ),
2352            BuildError::PluginOnReady { plugin, source } => {
2353                write!(f, "umbral: plugin `{plugin}`'s on_ready failed: {source}")
2354            }
2355            BuildError::TemplatesInit(err) => {
2356                write!(f, "umbral: templates engine failed to initialise: {err}")
2357            }
2358            BuildError::PluginDatabaseAlias { plugin, alias } => write!(
2359                f,
2360                "umbral: plugin `{plugin}` requested database alias `{alias}`, which isn't \
2361                 registered; call .database(\"{alias}\", pool) on the builder before .build()"
2362            ),
2363            BuildError::CrossDatabaseForeignKey {
2364                model,
2365                field,
2366                model_db,
2367                target_db,
2368            } => write!(
2369                f,
2370                "umbral: model `{model}` (database `{model_db}`) has a foreign key \
2371                 `{field}` to a model on database `{target_db}`. A FOREIGN KEY \
2372                 constraint can't span databases. Either route both models to the \
2373                 same database, or mark the field `#[umbral(db_constraint = false)]` \
2374                 to keep it a logical-only relation (joins / select_related still \
2375                 work; no physical constraint is emitted)."
2376            ),
2377            BuildError::DatabaseBackendMismatch {
2378                url_backend,
2379                pool_backend,
2380            } => write!(
2381                f,
2382                "umbral: settings.database_url names backend `{url_backend}`, but the \
2383                 default pool passed to .database(...) is a `{pool_backend}` pool. \
2384                 Either change UMBRAL_DATABASE_URL to match the pool, or open the pool \
2385                 against a URL whose scheme matches umbral::db::connect."
2386            ),
2387            BuildError::DuplicateStaticNamespace {
2388                namespace,
2389                first_plugin,
2390                second_plugin,
2391            } => write!(
2392                f,
2393                "umbral: plugins `{first_plugin}` and `{second_plugin}` both declare the static \
2394                 namespace `{namespace}` via static_dirs(); namespaces must be unique \
2395                 (they key the /static/<namespace>/ URL and the static_root/<namespace>/ \
2396                 collected-asset dir). Rename one plugin's namespace."
2397            ),
2398            BuildError::UngatedMutatingRoutes { routes } => write!(
2399                f,
2400                "umbral: deny_ungated_mutations() is set and {} app mutating route(s) have no \
2401                 recorded permission: [{}]. Gate each with the umbral-permissions \
2402                 `Routes::require_permission(...)` builder so the framework records the \
2403                 permission (a hand-applied `.layer(permission_required(...))` is NOT visible \
2404                 to this audit — prefer the builder). If a route is intentionally public, \
2405                 register it through a permission-aware builder or drop the strict flag.",
2406                routes.len(),
2407                routes.join(", ")
2408            ),
2409        }
2410    }
2411}
2412
2413impl std::error::Error for BuildError {}
2414
2415#[cfg(test)]
2416mod audit_tests {
2417    use super::ungated_mutating_routes;
2418    use crate::routes::RouteSpec;
2419
2420    fn spec(methods: Vec<&'static str>, path: &str, perm: Option<&str>) -> RouteSpec {
2421        RouteSpec {
2422            path: path.to_string(),
2423            methods,
2424            permission: perm.map(str::to_string),
2425        }
2426    }
2427
2428    #[test]
2429    fn flags_ungated_mutating_routes_only() {
2430        let specs = vec![
2431            spec(vec!["GET"], "/", None),                     // read → ignored
2432            spec(vec!["POST"], "/contact", None),             // ungated mutating → flagged
2433            spec(vec!["POST"], "/posts", Some("blog.add")),   // gated → ignored
2434            spec(vec!["DELETE"], "/posts/{id}", None),        // ungated mutating → flagged
2435            spec(vec!["GET", "POST"], "/api/comments", None), // has a mutating verb → flagged
2436        ];
2437        let flagged = ungated_mutating_routes(&specs);
2438        assert_eq!(
2439            flagged,
2440            vec![
2441                "POST /contact".to_string(),
2442                "DELETE /posts/{id}".to_string(),
2443                "GET/POST /api/comments".to_string(),
2444            ]
2445        );
2446    }
2447
2448    #[test]
2449    fn no_warning_when_all_mutating_routes_are_gated_or_read_only() {
2450        let specs = vec![
2451            spec(vec!["GET"], "/", None),
2452            spec(vec!["POST"], "/posts", Some("blog.add")),
2453        ];
2454        assert!(ungated_mutating_routes(&specs).is_empty());
2455    }
2456}
2457
2458#[cfg(test)]
2459mod drain_tests {
2460    use super::drain_after;
2461    use std::time::{Duration, Instant};
2462
2463    /// `drain_after` awaits its signal, flips the process to draining, then holds
2464    /// for the delay before resolving — the sequence that lets `/readyz` report
2465    /// 503 while the server keeps accepting during the drain window (Kikosi #5).
2466    ///
2467    /// One test, walked in sequence: the draining flag is a process-global that
2468    /// `begin_drain` only ever sets, so splitting the zero-delay and with-delay
2469    /// cases into separate concurrent tests would race on it. A ready signal
2470    /// (`async {}`) exercises the drain logic without delivering a real SIGTERM.
2471    #[tokio::test]
2472    async fn signals_draining_and_holds_for_the_delay() {
2473        assert!(
2474            !crate::shutdown::is_draining(),
2475            "draining must start false — nothing has signalled shutdown yet",
2476        );
2477
2478        // Zero delay: marks draining, does not sleep (the historical
2479        // instant-shutdown behaviour).
2480        let started = Instant::now();
2481        drain_after(async {}, Duration::ZERO).await;
2482        assert!(
2483            crate::shutdown::is_draining(),
2484            "the signal must mark the process draining so /readyz goes 503",
2485        );
2486        assert!(
2487            started.elapsed() < Duration::from_millis(50),
2488            "zero delay must not sleep; took {:?}",
2489            started.elapsed(),
2490        );
2491
2492        // A non-zero delay holds before resolving, even though the process is
2493        // already draining (begin_drain is idempotent; the hold still applies).
2494        let started = Instant::now();
2495        drain_after(async {}, Duration::from_millis(120)).await;
2496        assert!(
2497            started.elapsed() >= Duration::from_millis(100),
2498            "must hold for ~the drain delay before resolving; held {:?}",
2499            started.elapsed(),
2500        );
2501    }
2502}