eval: tool_calls_min expectations penalize efficient models

[chaliy]

Problem

Two eval tasks use tool_calls_min expectations that penalize models that solve tasks efficiently in fewer bash calls:

error_graceful_parse — expects tool_calls_min:2, Codex solves in 1

The task asks to fix broken JSON and print a field. GPT-5.3-Codex writes a single script that fixes the JSON and parses it in one invocation. The eval expects ≥2 tool calls (presumably: one to inspect, one to fix+parse), but there's no reason a model shouldn't solve this in one call.

complex_todo_app — expects tool_calls_min:3, Codex solves in 1

The task asks to build a CLI TODO app and demonstrate it. Codex creates the script and runs the demo in a single bash invocation. The eval expects ≥3 calls, but a single well-structured script accomplishes everything.

Impact

• GPT-5.3-Codex fails both tasks only due to tool_calls_min, scoring correctly on all other checks
• Anthropic models naturally use more turns and pass, but this doesn't mean fewer turns is wrong

Proposal

Remove tool_calls_min from both tasks, or reduce them to tool_calls_min:1 (just verify the model actually calls bash). The eval should test outcomes, not prescribe how many calls a model makes.

Eval Reference

GPT-5.3-Codex eval run (2026-02-27): crates/bashkit-eval/results/eval-openai-responses-gpt-5.3-codex-2026-02-27-055543.md

[chaliy]

This was already addressed in commit f7327c6 (PR #389) — tool_calls_min was removed from both error_graceful_parse and complex_todo_app task expectations. The eval results file still reflects the pre-fix run; a fresh eval will show both tasks passing.