fix(remote-control): honor retries while connecting

[shiny-code-bot]

Summary

• replace the unbounded reconnect command path with a capacity-one pending signal so duplicate requests coalesce
• allow a later remoteControl/reconnect call while status is connecting to schedule a fresh attempt instead of becoming a permanent no-op
• defer the retry until the active auth/enrollment/connection attempt reaches a committed boundary, preventing cancellation of side-effecting work
• suppress stale connected or errored status publication when a reconnect was queued concurrently
• preserve and assert the remote-control subscribe cursor across reconnects

Physical-device proof

Using the official ChatGPT iOS app 1.2026.188, the paired phone reopened thread 019f873b-32b4-73c3-9fb6-8b38e3420834 on the same persistent engine and completed both delayed and fresh turns:

• IOS_TO_MAC_ISSUE385_20260724T042959Z
• IOS_TO_MAC_TRACE385_20260724T112159Z

The second-controller thread/read returned the same thread with five turns, and the supervised engine remained healthy on 127.0.0.1:4766.

Validation

• just fmt
• git diff --check
• cargo test -p codex-app-server-transport reconnect_ — 12 passed
• cargo test -p codex-app-server remote_control_reconnect_ — 2 passed
• focused reconnect, replay, auth-recovery, and active-attempt tests pass
• full transport suite reached 130 passing tests; three parallel-only failures passed individually, while remote_control_waits_for_account_id_before_enrolling remains a pre-existing failure reproduced unchanged on clean origin/main
• independent breaking-change, change-size, context, testing, and async-concurrency reviews found no actionable issue in the final remote-control patch

Advances #385.