Pure critical-path estimates for Delegate launch admission.

Subagent estimates use ceil(task_count / max_threads) scheduling waves multiplied by the resolved uniform child budget. Workflow estimates sum the largest effective step budget in each planned dependency wave. Omitted step budgets use the runtime Subagent default, while both omitted and explicit budgets are capped by the normalized workflow timeout for the current estimate and wave_budgets_ms. suggested_timeout_ms is the least sufficient fixed point after that cap moves: it sums each wave's uncapped declared/default maximum, and can therefore exceed the current estimate only when the current workflow timeout binds. The estimate details carry the resolved caller horizon, whether that wait horizon was explicitly supplied rather than defaulted from the delegate timeout, the normalized declared workflow timeout, and whether that workflow timeout was explicitly declared before default injection, separately from their effective minimum. Rejection recovery chooses the wait-horizon knob when it was explicit and the delegate-timeout knob when it was derived. An explicit workflow timeout is combined with that caller knob only when it also binds. An omitted workflow timeout instead remains a delegate-derived ceiling: with an explicit wait horizon, recovery independently widens the wait and/or delegate timeout that binds; with a derived wait horizon, widening the delegate timeout alone moves both values. Maps from older callers that lack binding-source evidence retain the prior classification. Relaunching once at the suggestion is sufficient. Separately from admission, capped_step_budgets/4 names every explicitly declared step budget an explicitly declared workflow timeout caps, so the dry-run can predict the runtime closed_by_workflow_timeout cancellation instead of leaving it implicit in the gap between estimated_critical_path_ms and suggested_timeout_ms. That warning is purely advisory: it never moves would_reject, an admission verdict, or an exit code. Workflow wave maxima deliberately form a conservative batch-admission sum; a work-conserving scheduler can realize less wall time by starting newly unblocked work before a whole batch finishes. Admission still fails closed on that sum. Retries, provider jitter, and orchestration overhead are intentionally excluded.