core: side-channel frames for executor progress and task logs #590

The wire half of the parent issue: new frames, and a way to introduce them without forcing scheduler and executors to upgrade together.

Frames

Added to ExecutorMessage in crates/taskito-core/src/worker/protocol.rs:

Progress { job_id: String, progress: i32 },
TaskLog  { job_id: String, task_name: String, level: String, message: String, extra_len: Option<usize> }, // + payload

extra_len follows the result_len convention already in Success: None is no payload, Some(0) an empty one. publish is a TaskLog at level result, so it needs no frame of its own.

Added to SchedulerMessage::Job, both #[serde(default)] so an executor built before this ignores them:

disabled_middleware: Vec<String>,
metadata: Option<String>,

SchedulerMessage::HelloAck grows #[serde(default)] capabilities: Vec<String>.

Neither addition bumps PROTOCOL_VERSION. Bumping it would reject every mismatched peer outright — deliberate for a real wire break, wrong here, and it would couple scheduler and executor releases in the one place #546 exists to decouple.

Client

ExecutorClient grows an outbound side-channel the bindings can hand to their dispatchers — the sink a running task reports through. Requirements:

• Tasks never touch the shared Mutex<FrameWriter>; the sink is a bounded channel drained by the existing result thread, so a task in a progress loop cannot stall the writer or starve results.
• Progress coalesces to latest-per-job under backlog. Logs cannot (each is data), so the queue drops oldest and counts what it dropped.
• Nothing is emitted unless hello_ack advertised the capability. Dropping silently on an unadvertised scheduler is correct — the executor cannot know why, and failing the job over a progress call would be worse than the degradation this replaces.

Done when

executor_tests.rs covers, over MemoryTransport against a real RemoteDispatcher: a progress frame reaching the scheduler; a publish arriving as a TaskLog at level result; the job frame carrying disabled_middleware and metadata intact; a full outbound queue coalescing progress and dropping logs rather than blocking the task; and a scheduler advertising no capability receiving no such frames at all.

Parent issue carries the rationale.