Catch invented automation fields at the MCP tool boundary, not in the shared writer

[bdraco]

In the end to end runs through Home Assistant the model invented a field on its first upsert_automation call twice (switch_id, then switch on switch.turn_off) and recovered only after validate_config reported the esphome error. #2793 tried to refuse such trees in the shared writer by checking params against the catalog's field lists and was closed: the catalog is a lossy copy of rules esphome owns (measured against esphome 2026.9.0: emontx.send_command requires command but the catalog lists only id, and 54 registry entries could not be compared at all), and the visual editor auto-applies the whole parsed tree through the same writer, so any gap makes an existing valid automation impossible to edit.

Direction: esphome is the validator. The model must run validate_config after every write and repair from esphome's own error (replace with expected). Remaining work is on the MCP side only:

• Description: done in Add an upsert_automation MCP tool that lands an automation in the right place #2789 (e179780), the tool now says esphome checks the fields and to run validate_config after every write and repair with expected.
• Consider running esphome config inside upsert_automation and appending its result to the reply, so the model sees the real error in the same call; only if it fits Home Assistant's 10 second per call budget on a Pi, otherwise leave the two calls separate.

The shared automations/upsert writer stays validation free. Context: #2789 (the upsert tool), #2793 (closed).