fix(gateway): preserve pending update prompts across restarts

[teknium1]

Salvage of #18477 onto current main. Cherry-picked @simbam99's commit with authorship preserved — no follow-up fixes needed, cherry-pick applied cleanly.

Summary

Stops _watch_update_progress() from unlinking .update_prompt.json the moment it forwards the prompt. The marker now stays on disk until the user actually replies (or cancels via recognized slash command), so a gateway restart mid-prompt can recover the outstanding question.

Why

On current main: user runs /update on Telegram/Discord → update subprocess asks "Restore local changes? [Y/n]" → gateway forwards → gateway restarts before the user replies. The prompt file was deleted eagerly and _update_prompt_pending is in-memory only, so the user's y falls through as a normal chat message and the detached update subprocess hangs on .update_response until its 300s timeout fires.

With this fix, _schedule_update_notification_watch() at gateway startup (run.py:3314) already re-fires _watch_update_progress() whenever .update_pending.json is on disk — that watcher now finds .update_prompt.json still present and re-forwards it.

Changes

• gateway/run.py: remove eager prompt_path.unlink() after forwarding; add prompt_path.unlink(missing_ok=True) on the two sites that write .update_response (normal reply + recognized-slash-command cancel).
• Duplicate-send suppression within a single watcher still works — _update_prompt_pending[session_key] = True is untouched.
• tests/gateway/test_update_streaming.py: 5 new/updated cases covering restart-recovery, cleanup on normal reply, recognized-slash cancel, and unrecognized-slash response paths.

Verification

• Cherry-pick applied cleanly onto current origin/main.
• scripts/run_tests.sh tests/gateway/test_update_streaming.py -v → 21 passed including the new test_prompt_is_recovered_after_watcher_restart.
• Confirmed update-subprocess side (_gateway_prompt in hermes_cli/main.py:5200) only polls .update_response — never re-reads .update_prompt.json — so keeping the marker around has zero effect on the writer.

Closes #18477.