Faol hardening, and the supervisor is gone

Afternoon follow-up to this morning's migration.

The launchd gateway works now. The EX_CONFIG failures weren't a config problem: the plist pointed at /Volumes/Storage/Mac/agents/hermes/, and the install had moved into ~/.hermes. Reinstalled against the live path, it's been up since midday, multiplexing Telegram and cron ticks for both profiles.

Free fallback chain. Mirrored my default profile's fallback into faol's: OpenRouter thinkingmachines/inkling-small:free, key in the profile env. A forced-fallback turn answers in about five seconds. This mattered sooner than expected — the Haiku bucket on the shared subscription hit 429s this afternoon, and one of his Telegram turns spent six hours retrying before dying with max_retries_exhausted. Hermes' default policy is a 600-second sleep, three attempts, before the fallback chain is even consulted, so I set api_max_retries: 1 on the profile: one 429, straight over to the free model, turn completes in ~10s. He rides the fallback until the rate-limit window resets, then quietly returns to Haiku.

The tmux supervisor is deleted. Session killed, scripts and state removed. The heartbeat and the 3-minute mention watcher went with it — this morning's entry claimed that substrate survived the migration; it did, for about six hours. What remains is the cleaner architecture: the launchd gateway owns Telegram and cron, and faol's native cron self-wakes still fire through it. If I want a heartbeat back, it comes back as a gateway cron job, not a shell loop.