refactor(autopilot): canonicalize notification routing and add status diagnostics

[ahmedhesham6]

Description

Modernizes the autopilot notification routing system with canonical field names, per-schedule overrides, and observability improvements.

Changes Made

• Canonical target field — Renames chat_id → target as the primary notification destination field throughout config and runtime, with backward-compatible legacy aliases (notify_chat_id, schedule-level channel) that resolve in priority order
• Per-schedule notify_channel / notify_target — Adds schedule-level notification overrides replacing the legacy channel key, with conflict validation rejecting --target + --chat-id or --max-turns + --max-steps flag combinations
• max_turns per-schedule override — Replaces max_steps, propagated through apply_schedule_run_overrides into RunOverrides
• CheckTriggerOn default → Failure — Swaps from Success to Failure as the sensible default (trigger on check failures)
• NotificationRoute type — Renames DeliveryConfig → NotificationRoute with type alias for backward compat
• Notification status in autopilot status — New table column showing resolved route (e.g., slack:#ops) and readiness state
• Notification diagnostics in autopilot doctor — Detects missing channel/target, legacy key conflicts ([notifications].target vs chat_id, notify_target vs legacy aliases), and gateway reachability for delivery
• Live sandbox health — Doctor now fetches and displays sandbox health from the server health endpoint
• notification_route_config_warnings() — Warns when canonical fields override legacy aliases in the same config section
• Documentation updates — System prompt, channel setup guide, gateway README

Related Issues

Part of the autopilot notification routing modernization effort.

Testing

• ✅ All existing tests pass
• ✅ cargo fmt --check clean
• ✅ cargo clippy --all-targets -- -D warnings clean
• ✅ New tests: legacy compat, canonical priority, CRUD field preservation, conflict validation, doctor diagnostics, gateway target conversion, schedule max_turns override propagation

Breaking Changes

• CheckTriggerOn default changed from Success to Failure — schedules without explicit trigger_on will now trigger on non-zero exit codes instead of exit 0. This is a behavioral change but aligns with the more intuitive default.
• Legacy chat_id and schedule-level channel fields are still loaded but channel add now writes the canonical target field. Config files with both will receive warning diagnostics from autopilot doctor.