* Parametrize register_routes with the statics package

Lets extending servers reuse the statics mount and 404 handler
instead of copying them.

* Rename the factory UI flavor to enterprise

* Add register_plugin extension point

Lets extending servers register plugins that survive load_plugins, which
replaces the plugin list from the server config.

* Read the active Alembic config on each migration invocation

* Add Factory UI flavor and shared provider-aware login

* [Website] Update dstack Sky page

* [Website] Enable React Router transition flag

* [Website] Use thinner product selection outline

* [Website] Align product links and docs hover states

* [Website] Add product popup hover feedback

* [Docs] Open dstack Sky in a new tab

* [Website] Clarify dstack Sky hero description

* [Website] Swap dstack Sky hero and marketplace titles

* [Website] Refine dstack Sky marketplace heading

* [Website] Shorten dstack Sky hero description

* [Website] Align dstack Sky metadata and product copy

* [Website] Refine hero copy and Get started links

* [Website] Add dstack Factory landing page

* [Website] Trim contributor notes

---------

Co-authored-by: Andrey Cheptsov <[email protected]>

Upgrade the COS image used for these instance
types to the currently
[recommended](https://docs.cloud.google.com/compute/docs/gpus/gpudirect#create-high-edge-instances)
`cos-121-lts` family that comes with a GLIBC
version (2.37) compatible with `dstack-shim`.

* Fixed two `DetachedInstanceError` due to missing `RunModel.run_name`
  and `JobModel.job_name`.
* Fixed `Could not fetch server_info for worker ...: ReadError('')`
  logged as an error with a traceback, repeating until the worker is
  ready. An SSH tunnel to a port nobody listens on yet fails with
  `httpx.ReadError`, so this is the normal startup path. Both
  `_probe_http_worker()` and `_get_router_workers()` now catch
  `httpx.RequestError` and log it at the debug level.

Also:

* `_get_router_workers()` returns `None` instead of `[]` when the
  response cannot be read. `[]` means "the router has no workers", so
  an unexpected status, an unparsable body, or a missing `workers` key
  made the sync re-register every worker.
* Bare `except Exception` clauses are replaced with the exceptions
  actually expected, so a programming error is no longer reported as
  an unready replica. `sync_router_workers_for_run_model()` logs such
  errors itself instead of letting them escape to the pipeline worker,
  which leaves the sync row locked until the lock expires.
* `_get_worker()` takes the replica address instead of two pre-built
  URLs, and each probe builds the URL it registers the worker under.
  `_get_http_worker()`, `_get_grpc_worker()`, and
  `_discover_grpc_server_info()` are folded into loops over the
  connection modes and the runtime types to probe.

Fixes: #4300

See: kvesteri/sqlalchemy-utils#800

The sync only picked up router and worker jobs with
`JobModel.ready`, i.e. with all probes passed. A router probe
that depends on registered workers, such as `/health_generate`,
never passed, since workers are only registered once the router
is picked up, so the service deadlocked.

The sync now takes all running jobs, filtered by `job_num == 0`
instead. Workers are registered once the sync's own check
(`/server_info` or `GetServerInfo`) reports them ready, so
`probes` can no longer delay worker registration. As probes are
configured per service, not per replica group, such gating only
worked when the router and workers share the transport and API.

The `job_num == 0` condition is a no-op for now, as services
don't support multi-node replicas. Previously, it was implied by
`ready`, which is only set for the first node.

Fixes: #4310
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>

Co-authored-by: Andrey Cheptsov <[email protected]>

)

Co-authored-by: Jvst Me <[email protected]>

During a rolling deployment, the replacement router runs alongside
the old one, but the worker sync only picked the first running router
job, which is always the old one. The replacement had no workers
until the old router was scaled down and could not serve requests
routed to it.

The sync now reads the workers of every running router, probes the
workers once, and updates each router that is out of sync, so an
unreachable router no longer blocks the others. Router tunnels are no
longer held while workers are probed, which may take minutes: each
router is read first, and connected to again only if it needs an
update, re-reading its workers before applying the changes.

Also:

* `_Worker` is renamed back to `_TargetWorker`, its fields are typed
  after SMG's `WorkerSpec`, and `_add_worker_to_router()` sends it as
  the `POST /workers` body as is.

Fixes: #4315
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>

Since #4270, a router replica of a service with `model` and no
`probes` was probed with `GET /health` instead of the chat completions
request. The latter deadlocked: the router only answers it once
workers are registered, and the worker sync skipped routers that were
not ready. #4313 dropped the readiness check from the sync, so this
deadlock is gone.

`/health` passes before the router has any workers. In a rolling
deployment, the replacement router was considered ready, and the old
one was scaled down, before the replacement could serve requests. Now
that #4320 syncs workers with every running router, the replacement
passes the chat completions probe once it has workers.

Workers behind a router still get no default probe, as they may not
serve HTTP at all (gRPC workers).

With a Dynamo router, a router rolling deployment now waits for the
replacement indefinitely instead of scaling down the old router, as
workers stay attached to the old router's etcd/NATS. Such deployments
were broken before as well and are to be fixed separately.

Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>

* [Presets] Update system_prompt.md for router usage

* Resolve Review Comments

* Require router on cluster fleets with a CPU-only instance

---------

Co-authored-by: Bihan  Rana

Fixes #4318

Previously, all jobs displayed metrics for replica_num=0

* Apply DSTACK_SERVICE_CLIENT_TIMEOUT to gateway nginx and app

The server now passes the timeout to gateways on service and entrypoint
registration. The default is raised from 60 to 300 seconds to match the
previous nginx timeout.

* Fix docstrings

* Clean up tests

* Add Daytona backend

* Shorten Daytona volume documentation

* Validate Daytona API key permissions

Document required permissions and remove redundant Daytona tests.

* Pin gpuhunt 0.1.31

* Clarify Daytona volume sharing

* Clarify Daytona volume documentation

* Clean up Daytona docs, timeouts, and tests

---------

Co-authored-by: Andrey Cheptsov <[email protected]>

Starting October 8, Nebius is introducing Dynamic
Spot pricing for preemptible VMs. The price can
change for existing instances during their
lifetime.

To prevent the price from increasing above the one
`dstack` shows, `dstack` now caps the price at its
current value at provisioning time using a pricing
policy. If the price grows above the cap, the
instance is interrupted by Nebius. The actual
price may still drop below the one `dstack` shows.

Preemptible L40S is still sold at a flat rate and
is exempt from pricing policies.

Full support for Nebius Dynamic Spot pricing
(tracking the current price and showing it in
`dstack`, user-specified caps, etc.) may be added
later upon request.

Add an OCI backend setting that accepts a list of
shape names to allow provisioning in addition to
the standard supported instance families. Only
works for shapes included in `dstack`'s pricing
catalog (`gpuhunt`).

This setting can be used to test various shapes
that were not yet tested internally at `dstack`.

```yaml
experimental_instance_types: [BM.GPU.RTXPRO.8]
```