Point the kernel benchmark at a path that exists (#775)

[Rafael-SOWNet]

Fixes #775.

Both triggers filtered on "./Sources/AngouriMath/AngouriMath", which is not a path in this tree. The kernel is Sources/AngouriMath/ and its project file is Sources/AngouriMath/AngouriMath.csproj — there is no nested AngouriMath directory, and the pattern has no wildcard, so it could only ever match that one literal non-existent path.

The workflow last ran on 2026-01-02T14:24Z. The filter was re-added in f51c7e03 at 14:26Z — two minutes later. The January runs happened during the #640 CI work while the filter was temporarily absent, which is why the history looks alive up to that point and then stops dead.

Ninety-eight PRs have merged since. None were benchmarked, so a PR could have regressed Simplify by any factor with CI entirely green. These are the inter-version CPU and RAM benchmarks behind #529 and #500, and the only automated guard on not losing speed or memory on popular use cases.

The change

-      - "./Sources/AngouriMath/AngouriMath"
+      - "Sources/AngouriMath/**"
+      - "Sources/Tests/DotnetBenchmark/**"
+      - ".github/workflows/Benchmark.yml"

Two paths beyond the kernel, each for a reason:

• the benchmark harness — a change to what is measured is worth measuring;
• this file — which is also what makes the fix self-demonstrating. Kernel Benchmark has not run since 2026-01-02: its path filter targets a directory that does not exist #775 asks that whoever takes this "confirm the trigger fires rather than reasoning about the glob", and a PR that changes only a workflow file otherwise cannot. Because Benchmark.yml is now in its own path list, Kernel Benchmark should appear on this PR's checks — if it does not, the fix is wrong and you can see that without merging it.

workflow_dispatch added as well, so the benchmark can be run against a branch on demand. It could not be before.

The leading ./ was separately wrong — path filters are globs matched against repository-root-relative paths — though it made no difference while the target did not exist either way.

Two notes for review

The list is written out twice rather than shared through a YAML anchor. Anchors are not a supported part of the workflow syntax, and a benchmark that fails to parse is a worse outcome than the one being fixed.

Whether the pull_request leg should filter at all is still open, as #775 says: the job takes 9–11 minutes. This PR keeps the filter and makes it correct rather than deciding that question — but now that it works, dropping the filter is a one-line follow-up if the maintainers would rather every PR be benchmarked.

Checked

• the file parses, and both triggers resolve to the three paths above
• Sources/Tests/DotnetBenchmark builds clean on .NET 10, and CommonFunctionsInterVersion and RAMUsageTest — the two names the job passes on the command line — both still resolve in Program.cs

🤖 Generated with Claude Code

[Rafael-SOWNet]

Verified, rather than reasoned about, as the issue asks.

2026-08-07T17:54:09  completed/success  pull_request  fix/775-kernel-benchmark-path
2026-01-02T14:24:22  completed/success  pull_request  issue-620

Those are the two most recent runs of Kernel Benchmark. The first is this PR — the workflow triggered because .github/workflows/Benchmark.yml is in its own new path list, and it succeeded, which also establishes that the job itself still works after seven months of drift on .NET 10.

Run: https://github.com/asc-community/AngouriMath/actions/runs/31204448891