Refine AGENTS.md guidance on regression coverage and code comments - #30263

Summary

• Recommend microbenchmarks for regression coverage of hangs and pathological slowdowns. Agents keep trying to add mdtest coverage for these issues, which is very bad UX for humans running tests (you can't tell which test is hanging if the test "fails") and for agents, which often don't notice that the test suite has been running for 10 minutes and that this likely indicates that there's a problem.
• Drop the blanket reminders to test every significant change and get tests passing. Frontier models these days are, if anything, pretty over-eager in adding exhaustive regression coverage. I feel like these reminders are probably working against us these days, and may be causing agents to add redundant/unnecessary tests in lots of cases.
• Soften the language regarding comments. Most developers on Ruff/ty are using codex these days, which (I would argue) generally writes too few comments. That bullet was written at a time when most devs on Ruff/ty were using Claude, which often wrote very fatuous comments.