[Feat] Add opt-in launch criteria to custom automations

[roomote-roomote]

​Opened on behalf of @daniel-lxs. Follow up by mentioning @roomote-roomote, in the web UI, or in Telegram.

What changed

• Added optional plain-language launch criteria and typed runWhen conditions to custom automations in Settings, the REST API, and manage_custom_automations.
• Classified custom automation launch criteria as an internal-nightly experiment. Admins can see its control on Settings → Nightly Experiments only when the deployment opts into nightly experiments; launch evaluation remains off until the experiment itself is enabled. When off, the editor hides criteria, write endpoints reject them, MCP schemas omit the fields, and runtimes ignore saved conditions.
• Kept channel-based launch criteria independent. If a criteria-bearing Discord or Teams run was queued before the custom experiment was disabled, Fast restores its deferred destination root before replying; row-level enable/disable edits also omit saved criteria while disabled.
• Persisted launch criteria and per-run launch-gate results. This adds migration 0117_black_titania.sql; develop uses 0116_overrated_romulus.sql for task-run completion/cancellation indexes.

Why this change was made

Launch criteria change whether scheduled work begins, so their control and runtime behavior should be available only on deployments explicitly enabled for internal nightly experiments. Saved conditions and queued destination replies must remain safe when the experiment is off.

Impact

• The experiment defaults off. Existing saved conditions remain stored but inert while it is off, and channel-based launch criteria remain independent. Public automation documentation no longer points to the internal control.
• A bounded live smoke at PR head 8af03c33c7f86b8755fb16a131d8125ec637b1a1 confirmed the Nightly route is unavailable and criteria writes are rejected when the deployment opt-in is off. With the environment opt-in on but the experiment off, saved criteria survived four row toggles and a schedule update, and runtime ignored them. With the experiment on, real Jev returned a confident skip at 0.04 (saved privately and hidden from Results), a pass at 0.80 (continued to a shared result), and uncertain results at 0.65/0.66 (continued).
• A follow-up recovery check used synthetic pass and uncertain rows recorded through the PR-head result helper. The shared control was enqueued immediately and prepared in about four seconds. The private rows had no job at insertion; the scheduled five-minute recovery discovered and prepared both within two seconds of its next tick, in one attempt each. The PR-head listResultsCommand showed the shared control and excluded both private rows. For execution provenance, the result writer, shared enqueue, and Results query ran from a clean worktree at the PR head; the already-running background scheduler/worker performed recovery. Its queue and processor source is unchanged between that runtime revision and the PR head, so this does not claim the entire service stack ran at the PR head. All temporary rows, the disposable automation, queue jobs, and the three associated usage records were removed.

Screenshots

[roomote-community]

No new code issues found. See task

• Discord and Teams criteria-bearing automations can fail silently before their deferred destination root is created (packages/sdk/src/server/automations/custom-automations.ts:544).
• Failed Telegram DM topic creation falls back to an unthreaded report (packages/cloud-agents/src/server/fast-agent/fast-agent-service.ts:4508).
• show_widget can send a destination reply before launch criteria are evaluated (packages/cloud-agents/src/server/fast-agent/fast-agent-service.ts:2083).
• Disabling the experiment after a queued Discord or Teams run can skip its deferred root and post unthreaded (packages/cloud-agents/src/server/fast-agent/fast-agent-service.ts:2106).
• Disabling the experiment prevents toggling an automation that already has saved criteria (apps/web/src/components/settings/automations/CustomAutomationsSection.tsx:913).

_{Reviewed 6678563}

[roomote-roomote]

Fixed both outstanding findings in 7345263f: queued Discord/Teams runs now restore deferred roots when the experiment is disabled, and row-level automation toggles omit saved launch criteria while disabled so the update succeeds without clearing the saved rule.

The Roomote code review check reran on the pushed head and passed; the latest review summary reports no code issues, and both requested review threads are resolved. Focused web and Fast regression tests, affected typechecks, lint, and pre-push checks passed. The browser-proof capture for this follow-up timed out, so the PR body has no stale screenshots. See the task for the details.