Instrument test assemblies and raise test-class coverage to 100%

[Chris-Wolfgang]

Background

A fleet survey on 2026-08-14 found that ETL-SqlBulkCopy is the only repo measuring the coverage of its own test code. Every other repo instruments the src assembly only, so uncovered test code — dead helpers, unreachable branches, fakes nobody calls — has never been visible anywhere in the fleet.

Policy: test code is held to 100% line coverage. Test code that never executes has no purpose; it should be exercised or deleted. This is line coverage — branch coverage legitimately can't hit 100% (compiler-generated async state-machine branches), and the gate reads the line-coverage column only.

[ExcludeFromCodeCoverage] is a last resort, in this order:

1. Write a test.
2. If that's not possible, refactor the code to make it testable — without hurting performance.
3. Only if both fail, apply [ExcludeFromCodeCoverage], and justify it in review.

Status in this repo

Test assemblies are not instrumented. coverlet.runsettings exists but lacks <IncludeTestAssembly>true</IncludeTestAssembly>, so Summary.txt contains src rows only and current test-code coverage is unknown.

Related defect: the Stage 1 coverage gate mis-parses

.github/workflows/pr.yaml's Stage 1 gate anchors its filter on ^[^ ] to skip ReportGenerator's indented per-class rows, but the loop reads with a bare while read -r line. read strips leading IFS whitespace, so the indentation is gone before grep sees it and every class row matches — classes get gated as if they were projects.

This stayed invisible fleet-wide precisely because test assemblies aren't instrumented: with no class rows in Summary.txt, there was nothing to mis-match. It surfaced on ETL-SqlBulkCopy's 0.7.0 release, where the gate failed the release on a test class at 89.4% while both real assemblies were above 97%.

Fix (one line):

-          while read -r line; do
+          while IFS= read -r line; do

Stage 2's PowerShell parser is unaffected (Get-Content preserves indentation). Note .github/workflows/pr.yaml is a protected file, so that change needs a maintainer bypass merge.

Steps

• 1. Enable instrumentation — add to coverlet.runsettings:
  xml <IncludeTestAssembly>true</IncludeTestAssembly>
  Not a protected file, so this is an ordinary PR. Expect this PR to fail if test-code line coverage is below 90%. The Stage 1 gate already enforces CODECOV_MINIMUM against every assembly row, and enabling instrumentation adds the test assembly as a new row — so the threshold bites immediately, without step 4. That failure IS the measurement. (Confirmed in ETL-SqlBulkCopy's gate output: it checks Wolfgang.Etl.SqlBulkCopy.Tests.Unit as a module; it only passes because that assembly is at 99.5%.) Note release.yaml does NOT pass --settings coverlet.runsettings, so it will not instrument test assemblies even after this change — PR CI becomes stricter than release CI until step 5.
• 2. Measure — run the suite and read the per-class rows for the test assembly in CoverageReport/Summary.txt. Record the starting number here.
• 3. Raise every test class to 100% line coverage — following the escalation above. Expect genuinely dead test code; delete it rather than covering it for its own sake.
• 4. Fix the Stage 1 gate parse (see above) and extend it to enforce 100% on test-assembly classes while src assemblies stay at the existing 90% (CODECOV_MINIMUM).
• 5. Make release.yaml agree — it runs dotnet test without --settings coverlet.runsettings, so release-time coverage excludes different code than PR-time. Harmless today; once a 100% rule exists it means a release can fail a gate the PR passed.

Notes

• Do not soften the gate to unblock work. If enforcing 100% walls off PRs until the backlog is cleared, that is the intended behaviour.
• Steps 1–3 are per-repo. Step 4 is fleet-uniform and should land in repo-template first, then fan out.

---

Update: fleet sweep, 2026-08-14

A follow-up sweep found a second, larger reason test coverage is invisible: most test projects don't reference coverlet.collector, so dotnet test --collect:"XPlat Code Coverage" silently emits nothing for them — no error, no warning, just a missing coverage file.

Fleet totals: 83 test projects — 51 emit coverage, 31 emit none, 1 excluded by attribute. It is systematic by suite type, not random: Unit and Integration projects almost always have the collector; the specialty suites (AotSmoke, Fuzz, Concurrency, DocExamples, Snapshots, Allocation) were scaffolded without it across every repo that has them.

This repo — 1 of 1 test projects emit coverage

test project	emits coverage?
Wolfgang.Extensions.String.Tests.Unit	✅ yes

Revised step 1 — three parts, not one

1. Add coverlet.collector to every suite that should be measured.
2. Explicitly decide (and write down) whether AotSmoke / SourceLink-style suites stay uninstrumented.
3. Then add <IncludeTestAssembly>true</IncludeTestAssembly> — it only has an effect on projects that emit coverage at all.

Steps 2–5 in the original list are unchanged, but note the measurement in step 2 is only meaningful once step 1 is complete.

Scope decision, 2026-08-14

Two calls from Chris, narrowing this issue:

1. AotSmoke needs no coverage. Settled — AotSmoke suites stay uninstrumented deliberately. This is now a recorded decision, not an accident of scaffolding, and no longer an open question on this issue.

2. Phase 1 is unit tests only. Get the unit test projects to 100% first. The other specialty suites — Concurrency, DocExamples, Fuzz, Snapshots, Allocation, SourceLink — and the Integration suite are deferred to a later phase. Adding their collectors is explicitly not part of this issue's immediate work.

Deferred does not mean dismissed: the understated-src-coverage consequence noted above still holds while those suites emit nothing, so don't treat the post-phase-1 src number as the true figure.

Phase 1 scope for this repo

unit test project	status
Wolfgang.Extensions.String.Tests.Unit	✅ already emits coverage — only needs IncludeTestAssembly

Revised phase-1 checklist:

• Add coverlet.collector — not needed, the unit test project(s) already emit coverage.
• Add <IncludeTestAssembly>true</IncludeTestAssembly> to coverlet.runsettings. Expect the PR to fail if unit-test line coverage is under 90% — the gate enforces CODECOV_MINIMUM against every assembly row. That failure is the measurement.
• Raise every unit test class to 100% line coverage — write a test, else refactor to make it testable, and only then [ExcludeFromCodeCoverage].
• Gate change (100% on unit test assemblies, src stays at 90%) — lands in repo-template first, then fans out.