readme: state OpenAI's audit finding precisely

[mnkprs]

Why

§1 summarized OpenAI's Verified retirement as "at least 59.4% of those problems had tests that reject functionally correct submissions" and dated the post 23 February. Checked against the post itself: 59.4% is "material issues in test design and/or problem description" (35.5% narrow tests, 18.8% wide tests, 5.1% miscellaneous), the page shows no publication date, and the recommendation is broader than SWE-bench Pro alone.

What changed

• §1: the OpenAI bullet now carries the audit's actual method (138 hardest problems, ≥6 engineers per case) and the narrow/wide/miscellaneous breakdown, notes that this taxonomy is the report's two axes, softens the date to "In February 2026", and lists the full recommendation (Pro, private uncontaminated benchmarks, expert grading).
• §1: "left a 59% flaw rate on the hard subset" → "left material issues in 59% of the hard subset".