test(ci): cover the missing-upload-step branch with a synthetic job set (#303)

[phmatray]

Implements #303.

Closes #303.

Gives TestReportingTests a seam so its own assertions can be exercised against synthetic jobs. Test
infrastructure only — one file. No library code, no workflow files, no build/Build.cs change, no new
package.

Plan

• Task 1: Introduce the job-set seam and cover the absence path

What changed

TestReportingTests discovers its guard set once, statically, from the repository's real workflows.
Every real test-running job has the Publish: test-results upload step — that is what the suite
exists to keep true — so the five assertions' "the step is missing" branch could never execute in a
green run. #267 verified it by hand (delete the step from ci.yml, watch the tests name the job,
revert), which proved the behaviour once and could not prove it again.

• UploadOffenders(jobs, claim) takes the job set instead of reading JobsThatRunTests() itself.
  That parameter is the whole point: a test that cannot supply its own jobs cannot reach the branch.
• Each content claim is now defined once — RunsOnFailure,
  PointsAtTheDirectoryTheBuildFills, AvoidsThePathTheBuildNoLongerFills, ToleratesAMissingPath —
  referenced both by the [Fact] that holds it against the real workflows and by the coverage
  theories. A dictionary of re-typed copies would let the two drift apart silently.
• UploadMissing(jobs) names the presence check, so the test whose subject is "the step is there"
  no longer reads as jobs that fail an always-true claim.
• Three coverage tests: every claim reports a job with no step; every claim reports a step that is
  present and violates it; and a conforming step offends nothing.

Code review

A review raised 11 findings; all 11 addressed. One was decisive and two more were real gaps in the
coverage I had written — the first draft of this PR was half-built:

Finding	Why it mattered
No fixture had the step present but violating a claim	A UploadStepFails that ignored claim entirely would have passed every test, silently reducing all four content assertions to "the step exists" — the exact vacuity #303 was filed about, one branch over
Both fixtures were passed as single-element lists	An implementation returning the whole list whenever any job offended would have passed — and in production (3 jobs) that destroys the "names the offending job" property this file is built around
UploadClaims duplicated the predicates rather than being what the [Fact]s read, while its doc claimed the opposite guarantee	165 lines of separation between the copies; drift would have been invisible in both directions
The conforming fixture omitted name: test-results and pinned @v4	That name:/path: pair is the one discrimination the path claim exists to make, so the fixture could not exercise it. Now copied from ci.yml verbatim
A foreach over claims in one [Fact] aborts at the first failure	Now [Theory], so each claim reports independently — the form the sibling WorkflowSourceTests already uses

The rest were accuracy fixes, again mostly in my prose: a doc promising a guarantee nothing enforced,
"drives every claim over a synthetic job" (the delegate is never invoked on that path — ||
short-circuits), and "cannot fire in a normal run" said of a branch that fires exactly when the guard
does its job.

Verification

The coverage was mutation-tested, not assumed. Three mutations, each of which a reviewer identified
as previously shipping green:

Mutation	Before	After
UploadStepFails ignores claim entirely	passed	4 tests fail
Absence branch inverted (is { } step &&)	caught	4 tests fail
UploadOffenders returns the whole list if any job offends	passed	8 tests fail

All restored afterwards; the tree is clean.

• ./build.sh Test (the exact CI target) — Restore/Compile/Test all succeeded on the merged tree.
• dotnet test -c Release — 1601 passed, 1 skipped, 0 failed (FormCraft.UnitTests 876;
  ForMudBlazor 567; ForFluentUI 158).
• dev moved three times during the run. The last merge overlapped this file — The .html test report is promised but unenforced, so half the artifact can vanish silently #276 (PR build(ci): guard every promised report kind, not just the trx (#276) #325)
  also edited TestReportingTests.cs. Git merged it without conflict, and I checked the result was
  semantically clean rather than trusting that: The .html test report is promised but unenforced, so half the artifact can vanish silently #276's additions are confined to the Build.cs
  report-kind assertions, touch nothing in the upload region, and all four production call sites still
  route through the named delegates.

[phmatray]

Merge blocked — not by this PR. dev is red, and merging it in makes this PR's CI red too.

• This PR's own CI was green at 62225e4 (build-and-test, ubuntu-latest, lint-pr-title all passed).
• Syncing dev in (87d84d2) brought in the failure: CollectionItemShapeGuardTests.No_Suite_Should_Re_Declare_A_Collection_Item_Shape_The_Fixture_Provides, in FormCraft.ForMudBlazor.UnitTests — a project this PR does not touch.
• Reproduced on a pristine checkout of origin/dev with none of this branch's commits: Failed: 1, Passed: 582. dev has been red on every commit since 0ad0d49.

Cause and remedies are in #345 — a collision between #306 (added the guard) and #308 (added the model it rejects), each green on its own branch.

Holding this PR rather than merging over a red bar. Once dev is green, re-run /merge-pr #333: the branch is already synced, reviewed and verified, so it should land without further work.

[phmatray]

Unblocked. dev is green again — #346 (a68cf29) fixed the shape-guard collision that #345 tracked, and #345 is closed.

Re-synced this branch onto dev (8d23d45) and re-verified, including a gate that did not exist when this PR was written: #307 landed ci(build): enforce dotnet format, so the build now runs dotnet format --verify-no-changes over the whole solution.

• ./build.sh Test — 1625 passed, 1 skipped, 0 failed (884 / 584 / 158)
• ./build.sh Format (the new gate) — exit 0, no violations in this branch

Ready to land.