add WebSocketConfig::with_request_modifier (#165)

[bryan-ashby]

Part of #165: the request modifier, the second half of what we agreed there (the static config is #210).

Builds on #210, which builds on #189 and #191. Until those merge, the diff includes their commits; the change here is the last commit.

Changes

• WebSocketConfig::with_request_modifier takes an async callback:

  F: Fn(Request<()>) -> Fut + Send + Sync + 'static
  Fut: Future<Output = Result<Request<()>, E>> + Send + 'static
  E: Into<Box<dyn Error + Send + Sync>>

  The callback receives the validated request from build_handshake_request and returns the request to send. Use cases: a short-lived signed token (e.g. an AWS IoT custom-authorizer signature), or a presigned URI.
• It runs on every attempt. Set on the WebSocketConfig given to ConnectOptions::with_websocket_config, it runs before every MqttClient connection attempt, including automatic reconnects, because the client reads that config on each attempt (add ConnectOptions::with_websocket_config for MqttClient (#165) #210). So the values never go stale.
• It's bounded and fails cleanly. The callback runs inside the existing connect timeout. An error fails the attempt with MqttError::ConnectionError("WebSocket request modifier failed: …") before anything is dialed, so a reconnecting client retries as it would after any connect failure.
• What it returns is sent as is. The with_header checks don't cover headers the callback adds, and tungstenite still requires its handshake headers exactly once. The connection goes to the returned request's URI; the docs note that a callback pointing it at another host must also update Host.
• Types: the callback is stored as a public, opaque RequestModifier (Clone, with a Debug that prints RequestModifier(..)). WebSocketConfig gains a request_modifier field, and RequestModifierError is the boxed error type.
• The handshake future is boxed, so the connect futures that await it stay under clippy's large_futures limit.

Tests

tests/websocket_request_modifier.rs, against a loopback server:

Test	Checks
modifier_adds_headers_computed_at_connect_time	a header the callback computes arrives; the callback runs once; it sees the configured headers
modifier_can_replace_the_uri	a presigned-style query added by the callback reaches the server
modifier_error_fails_connect_before_dialing	ConnectionError with the callback's message; the server sees no connection
modifier_runs_within_the_connect_timeout	a callback that never finishes ends in MqttError::Timeout

In tests/client_websocket_config.rs, request_modifier_runs_again_on_reconnect uses #210's server, which closes each connection after CONNACK. It checks that the first connect carries x-token: token-1 and the automatic reconnect carries token-2.

All five fail if connect ignores the modifier. The two #210 tests in that file still pass, as expected.

Locally, on the pinned 1.98.0 toolchain, fmt, clippy (workspace, wasm, and the lib with no features or a single feature), cargo doc -D warnings and the doctests pass. The full suite passes apart from the same machine-only failures noted on #189 and #191. Fork CI: bryan-ashby#4.

🤖 Generated with Claude Code

[fabracht]

Thanks Bryan, this is merged now. The modifier is a good fit for the presigned URL and short-lived token cases.

We made a few changes before merging:

• The request the modifier returns is checked before dialing. It has to keep the URI scheme, exactly one of each handshake header the transport sets, none of the other reserved headers, and a Sec-WebSocket-Protocol that still offers mqtt. Otherwise the attempt fails with MqttError::Configuration. Before, a modifier could add Content-Length, send a second Host, drop mqtt, or change the URI scheme, and the transport dialed anyway.
• MqttClient no longer wraps WebSocket errors, so a modifier that hangs comes back as MqttError::Timeout, and a modifier error is reported once instead of nested inside "WebSocket connect failed".
• A modifier failure is logged at error level, so a token service failing on reconnects shows up at the default log level.
• MqttClient doesn't resolve the host of a ws:// or wss:// address anymore, since the handshake dials the request's host anyway. That means the address passed to connect can be a placeholder that the modifier replaces. Together with open one websocket connection per connection attempt #217, the modifier now runs once per attempt instead of once per resolved address.
• RequestModifier and RequestModifierError are exported from mqtt5::transport, and the CHANGELOG lists the new request_modifier field under Breaking.

There are tests for each of these. It'll go out in the next release.