[devmotion]

Fixes #868.

min/max on Dual were generated from DiffRules. Those rules choose the partials by value only, so on a tie in value the result depended on the argument order rather than on isless, which breaks ties on the partials since #695:

julia> x = ForwardDiff.Dual(48.0, 1.0);

julia> min(x, 48.0), min(48.0, x)  # before
(Dual{Nothing}(48.0,1.0), Dual{Nothing}(48.0,0.0))

julia> min(x, 48.0), min(48.0, x)  # after
(Dual{Nothing}(48.0,0.0), Dual{Nothing}(48.0,0.0))

No DiffRules rule can express this tie-breaking, since a rule only sees the values. So this PR stops generating min/max from DiffRules and defines them in terms of isless. Like min/max on floats, they propagate NaN values. Mixed Dual/Real arguments are promoted by Base's min(::Real, ::Real).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.06%. Comparing base (7c157f9) to head (1aeb58b).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #869      +/-   ##
==========================================
- Coverage   91.14%   91.06%   -0.08%     
==========================================
  Files          11       11              
  Lines        1084     1086       +2     
==========================================
+ Hits          988      989       +1     
- Misses         96       97       +1     

☔ 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.

[pkofod]

Looks right to me and behaves like #695. A case worth a test: under nesting, master's min can return
neither argument. With D(f) = t -> ForwardDiff.derivative(f, t) and
f(t) = min(3t - 2, t^2) at t = 1, master gives D(f)(1.0) == 3 (left branch)
and D(D(f))(1.0) == 2 (right branch). With this PR both come from t^2 (2 and 2),
in either argument order.

It may also deserve a line in the release notes, since derivatives at exact ties
change (d/dt max(t, 0) at 0 goes from 0 to 1), and tie gradients now depend on the
chunk size, as comparison-based code already does since #695.

[andreasnoack]

Not clear to me why GitHub isn't showing the "Update PR" button now that #875 has been merged.

[pkofod]

Not clear to me why GitHub isn't showing the "Update PR" button now that #875 has been merged.

you mean "Update branch"? Is "Always suggest updating pull request branches" on for this repo? Settings, General, Pull Requests.

[andreasnoack]

Yes. The setting is enabled.

[devmotion]

[andreasnoack]

I was trying out the new "Overview page" view, and apparently they haven't implemented the "Update PR" button yet.

[devmotion]

FYI the segfault on Julia 1.14 on windows is tracked upstream: JuliaLang/julia#64013