TEST cover Crescendo seeded media follow-ups

[Roman Lutz (romanlutz)]

Description

Crescendo already consumes original seed media after an accepted response and routes later turns using text or the latest generated media. This PR adds focused regression coverage for that contract, especially around cases where accidentally reattaching the original seed would reset an iterative edit.

The tests verify that:

• single-turn objective targets are rejected because Crescendo requires native multi-turn history;
• the original seed is attached exactly once during a normally accepted first turn;
• a content-filtered first attempt retries the same logical turn with the seed intact;
• after acceptance, later requests are text-only unless the latest accepted generated media is forwardable;
• adversarial prompt state, conversation branching, normalized and persisted history, scorer inputs, and terminal results remain consistent.

Production code is unchanged because the current routing behavior already satisfies these cases.

Tests and Documentation

• uv run pytest tests\unit\executor\attack\multi_turn\test_crescendo.py tests\unit\executor\attack\multi_turn\test_crescendo_resilience.py -q (91 passed)
• uv run ruff check tests\unit\executor\attack\multi_turn\test_crescendo_resilience.py
• uv run ruff format --check tests\unit\executor\attack\multi_turn\test_crescendo_resilience.py
• git diff --check origin/main...HEAD

Documentation was not changed because this PR only strengthens coverage for existing behavior. JupyText was not run because no documentation files changed.