Evals: a volume case for the communication rule, many done items and one buried problem #40

Before you file

• I searched the existing issues.
• I read evals/README.md.

Component

communication (rule)

Behaviour

The four cases under evals/rules/communication/ score 1.00 with and without the plugin (PR #31's table, Δ 0.00 on each), so they guard the rule's wording against regression but cannot show that the rule changes behaviour. The transcripts that motivated the rule showed the burying under volume: a hand-back with many done items and one problem in the middle, where a fresh session with one small task leads with the failure either way. The missing case is that volume case: a scaffold whose state gives the model a long list of things that went right and one thing that went wrong, graded on the problem being the first thing written.

Grading

A regex grader on the final message: the problem, named by what it is, appears before any done item. A second grader that the done items are present, so leaving them out is not a pass. Run with EVAL_ARGS="--ablation with-without" and expect a Δ above zero; if the baseline also leads with the problem at this volume, the case says so and the rule's effect stays unproved.

Prompt

A task on the calc fixture that produces, say, eight passing changes and one failing check or one carried finding, then asks for the hand-back. The exact shape is for the case's author; the point is volume with one buried problem.