The Prompter feature's integration tests were never executed by CI — the PR
checks run CodeQL and secret-scanning, not the pytest suite — so they carried
bugs that only surfaced once the feature was integrated and the full quality
gate ran against it:
- The project fixture set a `git_branch` field the model doesn't have (it is
`default_branch`) and omitted the required `assigned_cell` / `created_by`
columns, so every confirm-flow test errored at setup.
- The tests patched the Anthropic client's `messages.create` at the class
level, but the SDK exposes `messages` as a cached_property, so that target
can't be resolved. Route the call through a single `_create_message` seam on
PrompterService (behavior-identical) and patch that instead.
Also drop `confirm_draft` below the cyclomatic-complexity threshold by
extracting the override-merge, UUID-field, and enum-coercion helpers
(behavior-preserving; the prompter test suite passes unchanged).