MetadataSync: push scripts capture AI Prompt Run telemetry (transactional rows) alongside metadata

## What happens
`mj sync push` with `formatAsMigration` writes a `MetadataSync_Push_*.sql` script of everything the push executed. When a pushed record triggers an LLM-backed step — an `MJ: Queries` row's field extraction, an `MJ: AI Prompts` row's model run — the script also captures the resulting `spCreateAIPromptRun` / `spUpdateAIPromptRun` calls: 100 KB+ `Messages`/`Result` blobs with hardcoded PromptID / ModelID / VendorID that must exist in every target database.

Seen twice this week in Betty PRs (a query push carried ~600 lines of prompt-run rows; a prompt push carried two). Each had to be stripped by hand before commit, and one reviewer had to explain why a "metadata" migration contained conversation transcripts.

## Proposal
The push log should exclude transactional entities by default — at minimum `MJ: AI Prompt Runs`, `MJ: AI Agent Runs`, `MJ: AI Agent Run Steps`, `MJ: Action Execution Logs` — or, more generally, any entity not in the pushed directory set and not reached through a declared `relatedEntities` link. An `excludeEntitiesFromScript` list in `.mj-sync.json` would cover the rest.

Start at the `mj sync push` path when `formatAsMigration` generates `MetadataSync_Push_*.sql`, and trace how pushed entities and related entities enter the script. Compare the generated calls against the proposed exclusions and `.mj-sync.json` `excludeEntitiesFromScript`; done means transactional AI, agent, and action-log rows are omitted by default while declared pushed or related metadata remains.

Written by the indexing model from the issue text.