fix(desktop): reap stale gateway orphans on Desktop backend startup

[JoaoMarcos44]

Fixes #77276

Summary

Desktop restarts can leave a gateway process orphaned on hosts without a service supervisor (WSL, no-systemd Linux, some Windows setups), causing a duplicate gateway to run alongside the freshly-started one — which splits or duplicates inbound messages on connected platforms (QQ, Discord, etc.).

Root cause

Desktop tears down the previous backend two ways, neither of which reaches a detached gateway:

• POSIX: a plain SIGTERM to the backend's direct child only (apps/desktop/electron/backend-child.ts) — deliberately, since the backend itself isn't spawned detached and there's no process group to negative-PID-kill.
• Windows: a full process-tree kill (taskkill /T /F).

But the actual gateway daemon is spawned via hermes gateway restart, which detaches into its own session (start_new_session=True / Windows detach flags) specifically so it does survive backend churn intentionally for one-shot maintenance actions (backup, hermes-update, doctor, etc.). That same detachment means a plain SIGTERM never reaches it — and on hosts without a supervisor watching the gateway's lifecycle, the old instance just keeps running after the new backend starts and spawns its own.

I first considered patching _lifespan's shutdown to terminate() everything tracked in _ACTION_PROCS. Rejected: that dict also tracks legitimate one-shot actions that are supposed to outlive a backend restart (backup, hermes-update, doctor, skills-install...) — a blanket kill would corrupt in-flight backups/updates, trading one bug for a worse one.

Fix

hermes_cli/gateway.py already has _reap_unsupervised_gateway_orphans() — built and tested for the exact same failure mode on the CLI restart path (#51325, #75936), including the supports_systemd_services() gate so it's a no-op wherever a real supervisor already owns the gateway's lifecycle. This PR calls it once, on Desktop backend startup, before the new backend can spawn its own gateway — closing the Desktop-restart gap with already-proven logic instead of inventing new kill paths.

Test plan

• test_startup_reaps_gateway_orphans_when_desktop — sweep runs when HERMES_DESKTOP=1
• test_startup_skips_gateway_reap_when_not_desktop — server-mode backend does not run the desktop-only sweep
• tests/hermes_cli/test_web_server.py full suite: 134 passed, 5 skipped
• tests/hermes_cli/test_gateway.py full suite: 10 passed, 5 skipped (no regression on the reused reaper)

%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#8b0000', 'mainBkg': '#0a0204', 'primaryTextColor': '#ffccd5', 'primaryBorderColor': '#ff0038', 'lineColor': '#ff0038'}}}%%
graph TD
    A[🩸 Desktop Restart] -->|POSIX SIGTERM / Win taskkill /T| B[⚰️ Old Backend Dies]
    C[🔥 Detached Gateway Session] -->|start_new_session=True| D[👻 Survives As Orphan]
    B --> D
    D --> E{New Backend Boots}
    E -->|HERMES_DESKTOP=1| F[⚔️ _reap_stale_desktop_gateway_orphans]
    F -->|Reuses #51325/#75936 Reaper| G[🕯️ Orphan Exorcised Before Fresh Gateway Spawns]

Loading

Infographic :

[alt-glitch]

This was generated by AI during triage.

Related: #77721 also fixes #77276 using _reap_unsupervised_gateway_orphans() but invokes it from the restart path; this PR invokes it during desktop backend startup. These overlap but have different lifecycle coverage, so they need a maintainer consolidation choice.

[teknium1]

Resolved on main — and a credit correction is owed here: you submitted this startup-path reap on Aug 3, a day before #78312 (merged via #83406), making you the earliest submitter of this fix. Our duplicate sweep missed this PR at salvage time; apologies for that. Noting for the record that the design (reuse _reap_unsupervised_gateway_orphans in the Desktop lifespan under HERMES_DESKTOP=1, keeping the systemd no-op gate) was submitted here first.

The merged implementation is functionally equivalent to yours; #83542 additionally covered the restart path and shutdown termination. Thank you for the contribution and the clear analysis of the detached-session gap in your PR body.