The documentation samples run in CI, the wiki's and the website's

[Rafael-SOWNet]

Part of #1256: the documentation samples run in CI, the wiki's and the website's.

DocSamples compiles every code sample in the wiki against the library in the checkout, plus every sample on the website that an amcheck comment marks. It runs each one and checks it against the output its page states.

• It fails on a sample that doesn't compile, throws, or prints something its page doesn't say.
• A sample that doesn't finish fails nothing.
• It clones the wiki beside itself.
• In CI, AngouriMathSite is checked out beside the library and passed with --site.

It runs as its own job, since it builds a generated project per language. A change to the F# wrapper alone now triggers the workflow, because the F# samples compile against the wrapper. Measured.Commit(path) is back in the shared code, because this harness builds the library itself rather than referencing it.

What its first run found: against cd3b27b9, three wiki outputs that master no longer prints. Each is the same value in a different form:

page	the wiki said	master prints
Differentiation	16 + (cos(x) + a) * x	16 + x * (cos(x) + a)
Simplification	(1 + ln(x)) * x ^ x	x ^ x * (1 + ln(x))
Solvers	(-oo; -1/2 * sqrt(6)) \/ (sqrt(6) / 2; +oo) \/ (0; +oo)	(-oo; -1/2 * sqrt(6)) \/ (0; +oo)

The first two are #1635's factor order, and the third is a union that now absorbs an interval it contains. The wiki is corrected (AngouriMath.wiki d23ca13). With that, all 89 samples pass, 86 C# and 3 F#: 60 stated outputs verified and 0 failures.

The run before the wiki was corrected exited 1, with those three mismatches named. That is the harness's negative control, on real data.

After this, three of the harnesses #1256 lists are still outside CI:

• intbench needs the Rubi suite, which carries no licence;
• libcompare is a comparison, so nothing in it is a defect;
• egraph is a measurement of memory cost.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura