Add AGENTS.md

[Rafael-SOWNet]

Guidance for AI agents working on this repository. CONTRIBUTING.md covers how to build, test and open a PR; this covers how to do the mathematics, which is the part that actually goes wrong.

The substance:

• Mathematical correctness outranks backward compatibility. A published API that returns the wrong answer is a bug with users, not an asset to preserve. Same for notation: arcsinh is not a thing, which is what Refuse the arc- spellings of the inverse hyperbolic functions #687 was about.
• Consistency is the defect [Paper in progress] Next Step of Evolution: AngouriMath 2.0 #497 names, so when you fix something, ask what else is the same shape.
• An unevaluated result and a NaN are different claims. Unevaluated means "I could not settle this"; NaN means "this does not exist". Returning NaN for a limit you merely failed to compute is a wrong answer, not a graceful failure. Right answer > no answer > slow answer > wrong answer.
• Assert the mathematics, not the printed string — differentiate the integral back, substitute the root, subtract the identity. And reproduce on stock master before claiming something is broken.
• Aim at the hard problems, and measure against a corpus, reporting wrong/error/timeout counts and not just how many were solved. A change that solves one more and gets one wrong is a regression.
• Where to look things up: SymPy for reference implementations and cited papers, DLMF for special functions and branch cuts, Gruntz's thesis for limits, Bronstein for integration, ISSAC and JSC for what is new. Branch cuts get their own warning, since C99, .NET, Python and Mathematica do not all agree.
• Branch first, always, with a stated naming scheme, even with write access.
• Be concise, and keep wrong turns out of the code. Mistakes and retracted diagnoses belong in the issue tracker where they are searchable, not in commit messages or comments. What belongs in the code is the conclusion and the reason it holds.

It closes with the open issues worth starting on, grouped by depth, ending at the polynomial layer that most of the simplification and solving issues sit behind.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.59%. Comparing base (90c00a8) to head (8d0cad8).
⚠️ Report is 74 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #702      +/-   ##
==========================================
+ Coverage   80.99%   81.59%   +0.60%     
==========================================
  Files         155      157       +2     
  Lines       13687    13347     -340     
  Branches     1957     2198     +241     
==========================================
- Hits        11086    10891     -195     
+ Misses       1990     1849     -141     
+ Partials      611      607       -4     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.