Guard that every CI* target runs somewhere must walk the Nuke dependency graph, not grep workflow files

[jeremydmiller]

From the CI flakiness handoff (§4f), which proposed:

A guard that every Target CI* in build/CITargets.cs appears in some workflow, or is explicitly listed as manual.

The motivation is real and well earned: SlowTests existed, was maintained, was cited in issue write-ups, and ran in zero CI jobs — nothing anywhere reported that.

The catch

I prototyped the obvious version and it produces a false positive on day one:

$ comm -23 <targets defined> <targets named in .github/workflows/*.yml>
CIMessageRouting

CIMessageRouting appears in no workflow file, and it runs on every push — via build/build.cs:

Target CI => _ => _
    .DependsOn(CoreTests, CIMessageRouting);

and dotnet.yml runs ./build.sh ci.

A guard that cries wolf on its first run is how a signal channel gets ignored — which is the exact failure this whole reporting effort exists to undo, and is already called out in flakiness-report.sh as the reason NO_LEDGER_EXPECTED exists.

What it actually needs

Reachability, not name matching:

1. Collect every Target CI* declared in the build.
2. Collect the target names each workflow invokes (./build.sh <target>).
3. Expand those through DependsOn / Triggers transitively.
4. Report targets not in the transitive closure, minus an explicit allow-list of deliberately-manual ones (SlowTests is the precedent — it is workflow_dispatch-only by design).

Nuke can enumerate this itself; a hand-rolled grep of DependsOn would have the same brittleness as the naive version.

Priority

Low. This is a guard against a rare mistake, and it is only worth adding if it is accurate — an inaccurate version is worse than none.