fix(coverage): expose a de-duplicated first-party Cobertura aggregation (#815)

[drmoisan]

fix(coverage): expose a de-duplicated first-party Cobertura aggregation (#815)

Summary

• Adds Get-CoberturaFirstPartyCoverageSummary, which aggregates Cobertura line and branch counters across first-party packages while counting each (class, source line number) pair exactly once. The previous descendant-axis idiom counted a source line once per <line> element, so a line appearing under both the class rollup and one or more <methods> entries was counted repeatedly.
• Measured against a committed real Cobertura document, the defect doubles the branch counters exactly and inflates line counters by roughly a factor of two. The derived percentages barely move: branch percentage is bit-for-bit identical at 79.24%, line percentage differs by 0.01 point. This is a count defect, not a rate defect.
• Adds Format-CoberturaFirstPartyCoverageSummary (pure formatter) and Get-CoberturaFirstPartyCoverageReport (composition), so wiring the report into the coverage entry point costs exactly one line.
• Adds seven Pester tests, including a differential assertion that reproduces the old descendant-axis selection over the same fixture and asserts the corrected counts are strictly lower.
• Repairs a structurally invalid mocked Cobertura stub in two pre-existing test files. No assertion was weakened and no production validation was relaxed.

Why

spec.md records the root cause: the aggregation idiom selected .//line, a descendant axis that reaches <line> elements nested under <methods> as well as those in the class-level <lines> rollup. In the field-initializer-across-constructors shape confirmed at issue #670, one source line appears once in the class rollup and again under each constructor, so it is counted three or four times.

The corrected function delegates to the existing Get-CoberturaPackageLineSummary, which already de-duplicates by line number, rather than re-deriving the counting rule. That keeps one implementation of the de-duplication invariant instead of two.

What Changed

Core fix

• scripts/vscode/Invoke-MSTestWithCoverage.FirstParty.ps1 (new) — the three functions. Pure: no I/O, no mutation of the input document. Throws the same Cobertura XML does not contain a <packages> node. message as Get-CoberturaCoverageSummary for shape parity.
• scripts/vscode/Invoke-MSTestWithCoverage.Helpers.ps1 — one added line dot-sourcing the new file, matching the existing split precedent used by PackageRate.ps1 and Threshold.ps1. Helpers.ps1 sat at 469 lines against a 500-line ceiling, so a new file was the only placement that holds the ceiling.
• scripts/vscode/Invoke-MSTestWithCoverage.ps1 — one added line emitting the report, placed so the existing Done. Coverage artifact: line is retained.

Tests

• tests/scripts/vscode/Invoke-MSTestWithCoverage.FirstParty.Tests.ps1 (new) — seven tests over an in-memory here-string fixture. No temporary files.
• tests/scripts/vscode/Invoke-MSTest.RunSettings.Tests.ps1 and tests/scripts/vscode/Invoke-MSTestWithCoverage.AssemblyDiscovery.Tests.ps1 — the mocked ConvertTo-KoverageCoberturaXml returned <coverage line-rate="0.8" />, a stub with no <packages> node. See Risks below.

Docs

• spec.md — all 14 acceptance criteria checked off.
• plan.2026-09-08T23-49.md — all 48 tasks checked off.
• Evidence artifacts under the feature folder covering baselines, the fail-before/pass-after pair, the scope and threshold gates, and the QA loop.

Architecture / How It Fits Together

Invoke-MSTestWithCoverageMain writes the processed Cobertura document, asserts the line-coverage threshold, then calls Get-CoberturaFirstPartyCoverageReport on the same in-memory string and emits the returned line. That composition function casts to [xml], calls the aggregation, and hands the result to the pure formatter. Every computation and every emitted string lives in a unit-tested pure function; the entry point holds only the one-line call.

The allowlist is derived, not hard-coded: ProjectNames defaults to (Get-KoverageProjectAllowlist), evaluated at parameter binding.

Verification

Completed

PowerShell toolchain, run in order (format, then analyze, then test; type checking is not applicable per .claude/rules/powershell.md step 3 and is recorded rather than omitted):

Gate	Result
PoshQC format over both folders	ok, scoped git status --porcelain empty
PSScriptAnalyzer, scripts/vscode	16 findings, equal to the recorded pre-change baseline of 16; zero attributable to this change
PSScriptAnalyzer, tests/scripts/vscode	0 findings
Pester, tests/scripts/vscode	103 tests, 0 errors, 0 failures, against a baseline of 96
New module line coverage	96.97% against the 90% floor for newly added modules
Folder line coverage	78.56% to 79.47%

The loop restarted once from format after the stub repair described under Risks, then completed a clean single pass.

Fail-before/pass-after: the test file was authored before any production code existed and recorded red (FAILED=6, unresolved-command failures). After the fix it records PASSED=7 FAILED=0.

Differential counts over the fixture, executed rather than asserted from the plan:

Quantity	Descendant-axis (defect)	De-duplicated (fix)
LinesValid / LinesCovered	8 / 6	4 / 3
BranchesValid / BranchesCovered	16 / 10	8 / 4
LineRate	0.75	0.75

LineRate is identical under both computations, which is why the tests assert counts and not rates: a rate assertion would pass whether or not the fix is correct.

Corroboration against a committed real Cobertura document (66,158 de-duplicated lines):

Quantity	Descendant-axis	De-duplicated
LinesValid / LinesCovered	133,485 / 112,865	66,158 / 55,940
BranchesValid / BranchesCovered	33,624 / 26,642	16,812 / 13,321
Line percentage	84.55%	84.56%
Branch percentage	79.24%	79.24%

Recommended

• mcp__drm-copilot__run_poshqc_test with scan_folders set to tests/scripts/vscode.
• Run the coverage entry point end to end and confirm the added report line renders alongside the retained Done. Coverage artifact: line.

Backward Compatibility / Migration Notes

No breaking change. The three new functions are additive. The entry point gains one output line and retains its existing output. No public function signature changed, no file was renamed or removed, and no threshold, analyzer severity or policy requirement was altered.

Risks and Mitigations

The plan's coverage prediction for the wiring line was wrong, and it surfaced as a real test failure. Plan decision D3 asserted that no unit test reaches the post-processing block past the -NoExecute early return. Six pre-existing tests do reach it, mocking ConvertTo-KoverageCoberturaXml to return <coverage line-rate="0.8" />. The existing threshold assertion tolerates that stub because it reads only the root attribute; the new report does not, because it matches Get-CoberturaCoverageSummary and requires a <packages> node. The first test pass went red with 6 failures.

The repair made the stub structurally valid by adding an empty <packages /> at three sites in Invoke-MSTest.RunSettings.Tests.ps1 and one in Invoke-MSTestWithCoverage.AssemblyDiscovery.Tests.ps1, updating the one exact-match assertion quoting it in lockstep. Net zero lines in both files. The alternative — relaxing the new function's <packages> requirement — was rejected, because it would have diverged the new function from the existing Get-CoberturaCoverageSummary contract to accommodate a malformed test stub.

Both files are outside the Write Set spec.md names, which is a scope deviation and is called out here deliberately. Both sit under tests/scripts/vscode/, so the scope-boundary criterion is unaffected. The batch now touches 3 production and 3 test files, exactly at the .claude/rules/powershell.md cap. The largest file in either folder is 496 lines, under the 500 ceiling.

Consequence: the added wiring line is covered rather than uncovered, so the entry point's missed-line count is 10 before and 10 after.

Recorded findings, deliberately not actioned in this PR

• The corrected first-party line figure of 84.56% clears the 80% floor the script gate enforces and sits below the 85% floor stated in .claude/rules/general-unit-test.md and .claude/rules/quality-tiers.md. The gap exists under both computations (84.55% defective), so this correction neither created nor widened it.
• CLAUDE.md sets an 80% line floor while the two rules files set 85%, and the only automated script gate enforces 80. This divergence predates the issue. No threshold was lowered, weakened or deleted to make any gate pass.

Review Guide

1. scripts/vscode/Invoke-MSTestWithCoverage.FirstParty.ps1 — the aggregation, the rounding and zero-denominator fallback, and the delegation to Get-CoberturaPackageLineSummary.
2. tests/scripts/vscode/Invoke-MSTestWithCoverage.FirstParty.Tests.ps1 — the fixture, and the private differential helper that pins the old behaviour.
3. The two one-line production wiring changes.
4. The two stub repairs in the pre-existing test files (mechanical, net zero lines).
5. Evidence artifacts, if the numbers above need auditing.

Follow-ups

• Issue 828 (Bug: claude-md-cut3-names-uninvoked-coverage-command #828) records a separate documentation defect found during this work: CLAUDE.md section CUT3 step 4 names a vstest.console.exe ... /EnableCodeCoverage invocation that the repository deliberately never issues, because the built-in collector conflicts with the outer dotnet-coverage instrumentation. It is filed rather than fixed here: CLAUDE.md ownership between this repository and drm-copilot must be settled first, since everything under .claude/ other than agent-memory/ is pushed down with no templating.
• The 80-versus-85 percent line-floor divergence noted above has no issue of its own yet.

GitHub Auto-close

• Closes Bug: coverage-aggregation-double-counts-method-rows #815

🤖 Generated with Claude Code

https://claude.ai/code/session_01URyGon15RvDXcUEo96tvyh