Mastermind 3.3.1

Added

• Opt-in graph ranking for repo-maps with --ranking graph. Lexical ranking
  remains the default; graph ranking is an experiment without a quality claim.
• Offline source-quote binding audits for judge and producer assessments against
  pinned snapshots, including raw and rendered Markdown framing diagnostics.

Fixed

• Native review (review-task run, run-task --auto-review) is no longer lost
  when the model fences its JSON or adds a field: Claude Code validates the
  answer against the reviewer schema and the same model corrects it in the run.

Changed

• The native reviewer needs Claude Code 2.1.290 or later in the 2.1 series and
  records review
  policy v2 (Read, Grep, Glob and StructuredOutput). Completed reviews recorded
  under policy v1 stay valid.
• The reviewer has 30 turns by default, configurable per project with
  mastermind init --review-max-turns N. A run that stops at its turn or
  schema-retry limit records invocation_native_max_turns or
  invocation_native_structured_output_retries.

Install: npm install -g @xcraftmind/mastermind@3.3.1 or cargo install mmcg --version 3.3.1 --locked.