minimize: optional intra-message ddmin after turn-level shrink #34

Problem

minimize_genome (packages/mutiny_core/src/mutiny_core/minimize/__init__.py) implements classic ddmin over whole attacker messages (turns). Once a single long turn remains, content is not shrunk further. SYSTEM_DESIGN § minimize is message-oriented; for evidence UX, multi-sentence turns often stay verbose.

Repository evidence

• minimize_genome + _partition — turn subsets only
• tests/unit/test_minimize.py
• Regression artifacts store full conversation: list[str] (regress/__init__.py)

Why it matters

Smaller permanent regressions → clearer FAIL→PASS diffs and less secret surface in conversation text.

Proposed direction

After turn-level ddmin succeeds, optionally run a second-phase shrink on each remaining message (e.g. sentence/paragraph partitions or binary split of content) with the same re-exec gate (_interesting / same rule ids). Feature-flag or parameter intra_message: bool = False default off for behavior compatibility.

Acceptance criteria

• Default behavior unchanged when flag off
• When enabled, can reduce character length while still reproducing target violations
• Unit tests with a multi-sentence single-turn violating fixture
• No LLM in minimize path

Testing expectations

uv run pytest tests/unit/test_minimize.py -q

Scope / NOT included

• Tokenizers / model-based compression

Difficulty

L · core / advanced