Fix session recovery, SSH reverse binds and agent provider identity

[benfavre]

Delayed request responses could remove the newly selected detail sheet, and a rejected token left the app signed in and polling. Manage reads and writes now carry session ownership, each surface accepts only its latest read, and a current 401 returns to welcome while clearing account drafts and pending imports. Cloud sync commits fetched profiles only after that ownership check; refreshed identity also reaches Settings. A 403 remains a permission failure.

Reverse tunnels now use their configured remote bind address and explicitly cancel the exact listener on stop. Agent subscription help follows provider/host selection before the first run, and replies retain their original provider when a session switches from Claude to Codex. Legacy agent history still loads.

Validation:

• 857 unique workspace tests pass; four new regressions cover completion ordering, logout/re-login, 401/403 and durable provider attribution. UI tests and Clippy were rerun after the final lifecycle edits. Formatting, whitespace and Linux builds/checks pass.
• Native app screenshots reproduce the delayed-detail and expired-session failures, then verify recovery, draft cleanup and retained provider labels.
• All three tunnel types pass real HTTP traffic and native stop checks through a disposable loopback SSH server; the reverse listener receives explicit cancellation.
• Native Claude and Codex Agents runs complete over SSH using existing subscription credentials, without API-key environment variables. This uses an owned loopback target under the same OS account; fresh independent remote-host login is not claimed.
• Actual published v0.9.9 → v0.9.10 Linux update installs, verifies its hash and restarts successfully. No signing changes.
• Private X11 global palette/Dock shortcuts work with another test window focused; missing-tray hidden-start/close fallbacks pass. Physical tray/multi-monitor and native macOS/Windows interaction remain environment gaps.

Detailed evidence and limits: feature follow-up QA. Native rendered/click tests remain explicitly Red in the inventory rather than treating screenshots as automation.