MinimalPi baseline on sanitize-git-repo (current harness)

The A/B baseline for the RecursivePi read-heavy test (runs/smoke__qwen3.6-35b-a3b__20260710-130242/NOTES.md). Same model (qwen3.6-35b-a3b), same current harness, no recursive.

FAIL 0.0, agent 65 s / wall 101 s (faster than both RecursivePi runs, which paid 40–60 s of wasted child latency). Fails the same two verifier tests as RecursivePi — test_removal_of_secret_information (a 2nd HF token in a JSON git-diff string survives) and test_correct_replacement_of_secret_information (a clean line corrupted / exact-match failure) — while test_no_other_files_changed passes. No peg-native parser errors this run.

This is the clean isolation: the failure is a model-capability precision problem (faithful transcription of long random secrets, secret-vs-noise judgment, no over-replacement), identical with and without delegation. Recursive neither helps nor is the cause.