Revert #128: a benchmark baseline must be the bare model, no system prompt

[DietrichGebert]

Reverts #128.

A baseline arm is the control: the bare model with the task and nothing else. #128 gave it a system prompt ("Provide just one example for any given task, and no commentary or usage examples"), which:

• makes the "baseline (no skill)" label false (it now carries a hand-written instruction),
• desyncs the published numbers from the code: the single-shot table (baseline 518/693/256 LOC, "80-94% less code") was measured against the bare baseline and was never recomputed, so main shipped a baseline that would produce different numbers than the ones printed beside it,
• was undisclosed.

Any prompt on the baseline tilts the comparison ("write the minimum amount of code" would erase the gap; the opposite would inflate it). The fair control is no prompt at all. The rambling critique in #126 is already answered by the agentic benchmark, so the single-shot baseline does not need de-rambling.

🤖 Generated with Claude Code

[tosage05]

lol?