Collapse queue/steer into one gesture, add stop-and-reinject, stop discarding queued input on interrupt

[TheGreatAxios]

Summary

CL-5763: three defects in the mid-run send/steer/interrupt bindings.

• Plain Enter and Alt+Enter both queued mid-run and waited for a turn boundary — two gestures, one effect. Alt+Enter is now stop-and-reinject: it hard-stops the run immediately and restarts from the typed message, without waiting for a boundary. Plain Enter always queues to steer at the next boundary.
• Queued/steered/reinjected rows now say so on the row itself ([will steer next], [steering], [restarted here]) instead of leaving the operator to infer intent from a badge count.
• Interrupting (Ctrl+C) no longer discards queued or steered messages. It used to report interrupt — discarded N pending; it now reports interrupt — N queued message(s) will steer the next run, and those messages survive to steer whatever run starts next.

The chord check

Drove Shift+Enter through a live shell (tmux) rather than trusting the doc or the report. It does insert a newline, but only on a terminal that negotiates the kitty keyboard protocol (this app requests it via useKittyKeyboard); on a plain terminal Enter and Shift+Enter send the same bare \r, so the chord is a silent no-op there — which is what the operator saw. The shortcut list's existing wording ("Shift+Enter also works on terminals that report the modifier") already states that condition rather than promising it unconditionally, so it wasn't changed. Ctrl+Enter/Ctrl+J remain the newline chord guaranteed to work everywhere.

Sub-agent lanes on redirect

Both interrupt paths (Ctrl+C and Alt+Enter) close the underlying agent. That cascades: closing aborts the shared operation signal the task tool was given, which the tool forwards into the child agent's own controller, so an in-flight sub-agent dispatch is aborted along with the parent's turn and reports back as cancelled by the operator rather than running to completion detached. Decision: redirecting the parent stops the fleet it dispatched too — there is no path today to redirect the parent while leaving running lanes alone. Documented in docs/TUI.md.

Verified live (tmux)

• Queued a message mid-run with plain Enter — transcript showed [will steer next], then [steering] once delivered at the boundary.
• Queued a message, then interrupted before it drained — transcript showed interrupt — 1 queued message will steer the next run, badge stayed at 1, no discard. Sent a new idle message afterward and watched the surviving item steer into the new run as [steering].
• Typed a message and hit Alt+Enter while a sleep 8 tool call was in flight — saw stop — restarting from your message and [restarted here] ..., the run stopped immediately, and a new run started from the reinjected message without waiting for the sleep to finish.

Test plan

• bun run typecheck
• bun run build
• bun run test — 4277 pass, 0 fail
• keybindings.test.ts (the chord-through-a-live-shell suite) updated: Enter/Alt+Enter/Ctrl+C/Ctrl+G probes cover the new kinds; Ctrl+C probe also asserts a queued item survives a bridge-less local interrupt
• session-queue.test.ts, runtime-bridge.test.ts, shell.test.ts updated for the new interrupt/reinject semantics

[linear-code]

CL-5763

[TheGreatAxios]

Merged.

Rebased onto post-#430 main so steering and interrupt delivery share one model:

• Enter mid-run steers at the tool boundary (queue item, not a new send).
• Alt+Enter stop-and-reinjects: interrupt, then rebuild-safe deliver for the next prompt; rebuild failure surfaces a not-delivered notice rather than dropping the message.
• Ctrl+C keeps the queue and drains at interrupt via the Report fleet progress to the operator unprompted #430 deliver path (not boundary-only keep).

Transcript rows label pending intent; sub-agent abort-on-parent-redirect stays documented. CI snow flake (CL-5766) left alone — not this PR’s surface.