fleet: the supervisor verbs are tested and unreachable — wire them to the dashboard or delete them

[macanderson]

Fleet::pause_task / resume_task / stop_task exist and are tested in stella-fleet, and nothing calls them. Verified on main: every hit in stella-cli (fleet_cmd.rs:23, :522, :667, command_deck.rs:23) is a doc comment describing the verbs, not an invocation.

So the supervisor seam is real, tested, and unreachable by a user. The fleet dashboard already owns a live grid keyed by task id, which is exactly the surface these verbs were built for.

• Either wire the three verbs to the fleet dashboard's key handling, or delete them and the doc comments that promise them.

Wiring was deferred as a new user-facing feature with no witness test for interactive rendering — both true. But tested-and-unreachable code is the worst of both states: it carries maintenance cost and delivers nothing, while the docs tell readers the capability exists.

Acceptance: a supervisor can pause, resume and stop a task from the dashboard — or the verbs are gone.

_{Triaged out of #621 as worth doing.}

[macanderson]

Triage 2026-07-26 — premise confirmed, but the scope is ~4x what the body says

Confirmed unreachable. The three verbs are at stella-fleet/src/fleet.rs:781 (pause_task), :790 (resume_task), :801 (stop_task). A workspace-wide rg '\.pause_task\(|\.resume_task\(|\.stop_task\(' returns 12 hits, all inside #[cfg(test)] mod tests (lines 1168, 1172, 1585, 1587, 1608, 1609, 1623, 1624, 1625, 1634, 1635, 1636 — all past the mod tests boundary at :834). Zero production callers.

The receiving end is already wired, which is worth knowing before choosing delete: stella-cli/src/fleet_cmd.rs:663 destructures let WorkerControls { pause, stop } = controls; and races the stop line at :761/:801, with WatchGate gating the step loop. The seam is live and functional — only the sender is orphaned.

Why the dashboard genuinely cannot reach them: stella_tui::run_fleet_dashboard (stella-tui/src/fleet_dashboard.rs:566-571) takes inbound: UnboundedReceiver<FleetMsg> and done: oneshot::Receiver<()> — one-way only, no back-channel to Fleet. Its key handler (:634-646) binds only q/Esc/Ctrl-C, Up/k/Down/j, and s. The focused-row state (view, move_focus) does exist, so the surface is there; the plumbing is not.

Scope correction — the body says 4 doc-comment sites; there are ~17. stella-cli/src/fleet_cmd.rs has 5 (:23-24, :532-533, :677 — the body's :522/:667 have drifted by 10), stella-cli/src/command_deck.rs:23 has 1, plus 11 the body never counted inside stella-fleet itself: lib.rs:31-32, 49-51 (5) and fleet.rs:22-23, 71-72, 86, 142 (6).

So delete means retracting ~17 doc-comment promises across 4 files. Wire means adding an UnboundedSender<FleetCmd> to run_fleet_dashboard's signature, three key bindings on the already-focused row, and a Fleet handle reachable from fleet_cmd.rs:223.

This is the only issue in the #637–#645 group with no work in flight — draft PR #685 touches stella-fleet only for a one-line rustdoc fix in monitor.rs.