Feature: scheduled tasks run a pre-model precheck and skip the model turn when nothing changed #20258

Use case: An operator runs recurring checks (queue depth, CI state, an upstream API, "did any thread stop") and wants each check to run cheaply and wake an agent only when the state has changed. Today every fire of a scheduled task spends a model turn even when nothing changed.

Current behavior (as observed; not verified in source): each schedule fire starts a model turn. There is no step that runs before the model to decide whether the turn is needed. Related open issues: #19985 (a schedule cannot be bounded), #16818 (queued prompts accumulate while threads are blocked), #17889 (a failed scheduled run is recorded as succeeded).

Requested behavior: an optional precheck that runs a command or HTTP request before the model. If the precheck reports no change, the run is recorded as skipped with a reason and no model call is made.

API sketch:

schedule_task({ ..., precheck: { kind: "command", command: ["<check-script>"], timeoutSec: 30 } })
schedule_task({ ..., precheck: { kind: "webhook", url: "<endpoint>", changeFrom: "body-hash" } })
Precheck result { "changed": false } -> run status "skipped_precheck"
                { "changed": true }  -> model turn runs, precheck output in the prompt