fix(desktop): survive and resume turns interrupted mid-stream

[OutThisLife]

Summary

A turn interrupted by a disconnect, a renderer crash, or a full app/machine death used to leave the desktop with nothing to show for it: the gateway cleared its in-flight state while emitting terminal frames the detached client never received, nothing on the renderer side survived a process death, and the interrupted turn never finished. This closes that hole from both ends — the turn's progress survives, a failure renders as a failure, and a fresh interruption picks itself back up.

• The gateway keeps failed turns replayable. A turn that ends in error retains a compact inflight snapshot (user prompt, partial text, error, recoverable) until the next turn starts, on all three loss paths: returned-error results, turn exceptions (which now close with the same status:"error" message.complete frame as returned errors, instead of a bare error event), and agent-init failures.
• The desktop journals the running turn's visible tail — tool calls included — to localStorage, throttled off the delta-flush hot path, bounded, and cleared the moment the turn settles. Session resume folds it back onto the restored transcript. When the backend also has a live text-only snapshot for the same turn, the journal overlays its richer structure onto that row (longer text wins, base row id kept so live deltas keep landing) instead of treating the turn as caught up.
• Terminal error frames render as failures. The frame's structured error/partial fields drive the failed-bubble state instead of a text regex that missed the gateway's Error: <detail> texts. Partial output stays visible, and a retained failure replayed through resume renders as failed rather than as a healthy reply.
• Crash-interrupted turns continue themselves on resume. A turn writes a durable marker when it starts running and retires it the instant its outcome reaches the client — success, handled error, and interrupt all retire it — so a surviving marker is positive proof the process died mid-turn. Retirement happens at the terminal frame rather than at the end of the turn thread, because post-turn work (titles, memory sync, goal hooks) keeps running for a second or more after the reply lands, and quitting inside that window would otherwise look like a crash and re-run a finished turn. session.resume re-submits a fresh interruption with the original prompt in a recovery note, streamed live to the client that just resumed and shown as a compact "resumed interrupted turn" event row. Stale interruptions just leave the recovered partial in place for a manual continue. Config lives under desktop.auto_continue (enabled, freshness_minutes 15, max_attempts 2) — the same freshness window plus crash-loop breaker the messaging gateway's restart auto-resume already uses.

Supersedes #44339 and #46837; thanks @OmarB97 and @Reza2kn, both credited via Co-authored-by. Rebuilding onto current main also fixes the review defects on each: the backend-first merge ordering that dropped locally recorded tool progress, and the returned-error path that was never routed through retention.

Test plan

• scripts/run_tests.sh tests/tui_gateway/ tests/test_tui_gateway_server.py tests/test_tui_gateway_queue_on_busy.py tests/test_tui_gateway_ws.py tests/hermes_cli/test_config*.py — 1154 passed, including 8 new retention tests and 19 new auto-continue tests
• apps/desktop: full vitest suite — 2873 passed, including journal persist/merge, terminal error frames, and the auto-continue timeline row
• npx playwright test e2e/image-attachment-resume.spec.ts — the real-gateway seeding spec, which caught the marker-retirement window above
• npm run typecheck across all three desktop tsconfigs, plus eslint and ruff clean

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 082bd17}

all good!