Dispatch queued messages on the real turn boundary

[TheGreatAxios]

Summary

• Drain the mid-run message queue on inference.done, the only turn boundary the reactor guarantees once per cycle, instead of reactor.done, which fires once at agent shutdown and left queued messages stuck forever on any text-only reply
• Settle the turn's isProcessing on that same inference.done (once no tool calls are still outstanding, the same criterion connector.reply already uses) — a workflow/goal-governor cycle that keeps self-continuing may never emit connector.reply, and without this both run and the phase ramp stayed stuck ("working" forever, every later Enter routed to the queue instead of sending) once nothing else arrived
• Replace the queued-item transcript row's internal state-transition text (queue +1 → pending 1, with a duplicated queue label) with the queued message's own text and attachments, so depth is shown once (the notice row) and each item is identifiable

Verification

• bun run typecheck, bun run build, and bun test all pass — 4371 pass, 0 real failures. One local-only false positive: landing.test.ts's "no titlebar" assertion fails in this checkout because the worktree directory itself is named cl-5563-queued-messages-never-dispatch-and-the-queue-readout-is, which contains the literal substring "queue" and leaks into the rendered title bar the test greps for — unrelated to this change and not reproducible from any other checkout path
• Regression test reproducing the stuck queue on a tool-less turn (inference.start → inference.text.delta → inference.done, no tool calls, no connector.reply) — fails against the old reactor.done-only drain, passes with the fix
• Regression test reproducing the goal-governor case: after that same tool-less turn, both run and turnPhase return to idle, and a fresh Enter dispatches via sendImmediate instead of enqueuing into a queue nothing will ever drain
• Regression test (turn-state.test.ts) confirming inference.done settles the turn when no tool calls are outstanding, and does not settle when one still is — a real multi-cycle tool turn is never falsely marked idle mid-turn
• Ran the fix through the OpenTUI test harness end-to-end: queuing a message with an attached image during a busy turn, then completing the turn with a plain-text reply (no tool calls), dispatches the queued item automatically and shows queue 1 once in the notice row with the message text and [1 image attached: ...] on the same row, no +1 → or duplicated queue label
• Cancel-last (Ctrl+X) is intentionally not in this PR — it needs its own transcript-row handling and its own review, and is now tracked in CL-5572

Closes CL-5563

[linear-code]

CL-5563