fix(coverage): dedupe Cobertura line and branch arithmetic (#441, #478)

.claude/agent-memory/feature-review/project_441-review-residuals-and-494-handoff.md

@@ -0,0 +1,16 @@
+---
+name: 441-review-residuals-and-494-handoff
+description: "#441/#478 review (epic build-ci-coverage-gate-fidelity wave 0): PASS 0 blocking; residuals NF-1 (uncovered max-hits line 220) + NF-2 (poshqc zero-coverage artifact); corrected repo line-rate 85.0317% handoff owned by #494"
+metadata:
+  type: project
+---
+
+Review of `bug/cobertura-coverage-arithmetic-441` (2026-08-10T23-35, base edf3d34c): PASS, 0 blocking, all 20 spec ACs independently verified. Facts later epic-sibling reviews (#457 wave 1, #494 wave 2, #512) will need:
+
+- **Threshold margin handoff:** the corrected package-filtered line rate for the #424 sample is **62345 valid / 53013 covered = 0.850317 (85.0317%)** vs the uniform 85% floor — a 0.03 pp margin. Pre-fix inflated figure was 0.856453. #494 owns any threshold decision; #441 changed no threshold (verified empty diff on CLAUDE.md/.claude/rules/coverage.config). If #494's review sees a threshold edit, this is its provenance.
+- **Coverage-arithmetic oracle now available:** post-fix `Get-CoberturaCoverageSummary` reproduces a raw dotnet-coverage document's own root attributes exactly (79957/56124/23109/13472 on the #424 baseline). Any future review can use generator parity as a cheap correctness probe.
+- **Residual NF-1:** `Helpers.ps1:220` (`$existing.Hits = $hits`, the later-entry-larger-hits dedup update) has no covering test; recommended fixture rides a later change (#529/#530 work). If a later diff touches `Get-CoberturaClassLineSummary`, check whether the gap got closed.
+- **Residual NF-2:** filed as a recommendation only — the `run_poshqc_test` zero-coverage capture defect (see [[poshqc-bundled-coverage-artifact-reads-zero]]) was NOT promoted to an issue by the reviewer (no promotion tools in-session); the orchestrator was asked to file it or fold into #512. Verify it did not evaporate.
+- Follow-ups #529 (package rates stale), #530 (merged class keeps only primary `<methods>`), #531 (discovery lacks `\.claude\` exclusion — same defect as the local-vstest worktree pollution memory), #532 (wrong agent-memory generalization) are OPEN and deliberately unfixed in #441.
+
+**How to apply:** when reviewing #457/#494/#512, read `<441-FEATURE>/evidence/other/threshold-handoff-494.2026-08-10T23-15.md` before adjudicating any threshold or coverage-figure claim, and expect coverage baselines recaptured after b52874d6 to be non-comparable with pre-fix history (denominators shrank).

.claude/agent-memory/feature-review/project_poshqc-bundled-coverage-artifact-reads-zero.md

@@ -0,0 +1,12 @@
+---
+name: poshqc-bundled-coverage-artifact-reads-zero
+description: "run_poshqc_test writes artifacts/pester/powershell-coverage.xml with ZERO covered lines repo-wide (invalid capture); it is the hook's canonical PowerShell path and forces a FAIL row whenever PowerShell is enumerated"
+metadata:
+  type: project
+---
+
+The bundled MCP tool `mcp__drm-copilot__run_poshqc_test` writes producer output to `artifacts/pester/` (`pester-junit.xml`, `powershell-coverage.xml`, `powershell-coverage.koverage.xml`; gitignored via `.gitignore:57`). Observed in the #441 review (2026-08-10): the JaCoCo `powershell-coverage.xml` recorded **0 covered / 16075 missed lines for every file in the repo**, including a file whose committed direct-Invoke-Pester JaCoCo at the same head read 90.59%. The capture is an invalid instrument.
+
+**Why:** `artifacts/pester/powershell-coverage.xml` is exactly the path `validate-feature-review-coverage.ps1` reads for PowerShell repo-wide coverage (`Get-JacocoRepoCoverage`). A 0% reading is below the 85% floor, so once PowerShell is enumerated from `pr_context.summary.txt` the hook REQUIRES a FAIL verdict on a PowerShell coverage row. Pester JaCoCo emits no BRANCH counters, so the branch check returns null and skips (no unconditional block, unlike the C# #328/#398 cases).
+
+**How to apply:** In any TaskMaster PowerShell feature review: (1) check the artifact's actual counters before writing rows; (2) if it reads 0, write the repo-wide row as an honest FAIL against the artifact, disposition non-blocking as a pre-existing tool measurement defect, and adjudicate the real floors from the committed feature-evidence direct-Pester JaCoCo (changed-file, new-code, changed-line regression); (3) recommend filing a tooling follow-up. Executors document the MCP tool's payloads as "non-probative" — the direct `Invoke-Pester` runs are the probative evidence. Also note the #441 executor CORRECTION: an empty `git status --porcelain` does NOT prove the tool wrote nothing, because `artifacts/` is gitignored wholesale. Related: [[coverage-hook-skips-when-no-pr-context-summary]], [[project_stale-untracked-coverage-xml-leftover-false-block]].

docs/features/active/2026-08-10-cobertura-coverage-arithmetic-441/code-review.2026-08-10T23-35.md

@@ -0,0 +1,41 @@
+# Code Review — 2026-08-10-cobertura-coverage-arithmetic-441
+
+- **Timestamp:** 2026-08-10T23-35
+- **Reviewer:** feature-review agent
+- **Branch:** `bug/cobertura-coverage-arithmetic-441` at `3b8d43fb` vs base `edf3d34c`
+- **Files reviewed:** `scripts/vscode/Invoke-MSTestWithCoverage.Helpers.ps1` (357 -> 455 lines), `tests/scripts/vscode/Invoke-MSTestWithCoverage.Helpers.Tests.ps1` (222 -> 468 lines, pure addition)
+
+## Executive Summary
+
+The fix is well-scoped and correctly targeted. The one defective descendant-axis selection (`$cls.SelectNodes('.//lines/line')`, old line 122) is replaced by delegation to a new pure helper, `Get-CoberturaClassLineSummary`, which enumerates the class-level rollup and the method-level view over the child axis, keys by line number, and resolves duplicates by max hits / branch-if-either / larger-denominator condition-coverage precedence. The #478 blended-denominator defect is fixed by replacing the synthetic-document delegation with a direct helper call on the merged class node. The protected union builder is byte-identical to the base. Test additions are disciplined: fail-before was demonstrated with count-based (not rate-based) assertions, fixtures are pure in-memory here-strings, and the helper's precedence branches each have a dedicated unit test. No blocking defects were found. Two Minor and three Informational findings are recorded; the substantive one is a single untested statement (the max-hits update at line 220).
+
+## Findings Table
+
+| Severity | File | Location | Finding | Recommendation | Rationale | Evidence |
+| --- | --- | --- | --- | --- | --- | --- |
+| Minor | scripts/vscode/Invoke-MSTestWithCoverage.Helpers.ps1 | line 220 | `$existing.Hits = $hits` (the max-hits update when a later duplicate carries strictly larger hits) is executed by no test; every existing dedup fixture presents the larger or equal hits value first, so the update branch is dead in the suite | Add one fixture where the class-level rollup records `hits="0"` and a method-level entry records `hits="1"` for the same line number, asserting `CoveredLines = 1` | The helper's documented contract says duplicates resolve by max(hits); the contract's distinguishing case (later entry larger) is currently unpinned, so a future regression to first-entry-wins would pass the suite | Committed JaCoCo `evidence/qa-gates/pester-coverage-final.2026-08-10T23-10.xml` shows line 220 `mi=1 ci=0`; reviewer re-summed; executor recorded the same gap in `evidence/qa-gates/coverage-delta.2026-08-10T23-10.md` |
+| Minor | (tooling, outside diff) | artifacts/pester/powershell-coverage.xml | The bundled `run_poshqc_test` coverage capture records zero covered lines for every file in the repository, contradicted by direct-Pester evidence at the same head | File a follow-up tooling issue through the promotion lifecycle (or fold into #512); do not rely on this artifact for any coverage gate until fixed | An artifact that reads 0.00% for a file with a passing 19-test suite is an invalid instrument; it is also the canonical path the feature-review coverage hook reads for PowerShell | Reviewer parse of the artifact (LINE counters: 16075 missed, 0 covered); § 2.5 of the policy audit |
+| Info | docs/.../evidence/qa-gates/poshqc-analyze.2026-08-10T23-10.md | test-file section | Prose says "The 167 added lines introduced no analyzer debt"; the final test-file delta is 246 added lines (167 fixtures + 79 helper tests) | None required; correct if the artifact is ever revised | The finding-set arithmetic (0 findings before and after) is correct and was independently re-verified at HEAD; only the narrative line-count is stale | Reviewer `Invoke-ScriptAnalyzer` run; `git diff --numstat` |
+| Info | scripts/vscode/Invoke-MSTestWithCoverage.Helpers.ps1 | lines 368-371 | Merged-class rate expressions duplicate the rounding/fallback logic of `Get-CoberturaCoverageSummary` instead of sharing a second helper | None; accepted design | In-code comment records the rationale: the spec constrains the change to exactly one new helper, and existing assertions depend on identical `[math]::Round(...,6)` rounding and the `'0'` zero-denominator fallback | Diff hunk at old lines 267-279; test `line-rate | Should -Be '1'` |
+| Info | scripts/vscode/Invoke-MSTestWithCoverage.Helpers.ps1 | lines 251-259 | `Get-CoberturaClassLineSummary` returns its internal `$lineMap` (with live `Node` references into the caller's XML document) as part of its output object | None required now; consider documenting that `LineMap` entries are mutable views if a third consumer is ever added | Both current consumers only read the aggregate counters or (in the pre-existing merge path pattern) intentionally use nodes; the exposure is deliberate per the helper's `.OUTPUTS` help, and the helper mutates nothing itself | Source read; both call sites inspected |
+
+## Design and Correctness Assessment
+
+- **Root-cause fidelity.** The change edits exactly the defective site and its delegation path, and nothing else. The issue text's line references were wrong (declaration lines, not selections); the spec's corrected site table is what the diff implements. Reviewer verification: zero descendant-axis `.//lines/line` selections remain; the child-axis union builder is byte-identical (`cmp` old 217-268 vs new 311-362).
+- **Arithmetic correctness.** Independently reproduced on three inputs: the raw generator document (79957/56124/23109/13472, exactly matching the generator's own root attributes), the pre-change defective figures (161086/113219/46218/26944 — branches exactly 2x, lines slightly less than 2x due to non-uniform duplication, both consistent with the two-view duplication model), and the package-filtered A/B (62345/53013/0.850317).
+- **Precedence rule reuse.** The condition-coverage precedence (larger Total, tie-broken by larger Covered) reuses `Get-CoberturaLineConditionCoverageParts` rather than re-deriving the regex, and matches the precedence already used by the merge path — one rule, one implementation.
+- **StrictMode safety.** The switch from property access to `GetAttribute` is deliberate and commented (`Set-StrictMode -Version Latest` makes missing XML attribute property access throw). Behavior note: `GetAttribute` on an absent `hits` attribute returns an empty string and `[int]''` throws — fail-fast is acceptable here because Cobertura `<line>` elements always carry `number` and `hits`, and the pre-existing code had the same effective contract.
+- **Purity.** The helper performs no I/O and does not mutate the source document; `TotalLines` derives from the dedup map count, not the raw node count. Enumeration order (rollup first, then methods) combined with the update rules makes the result order-insensitive for hits and branch flags; the condition-coverage tie-break (strictly-greater Covered) is order-sensitive only between entries with identical Total and identical Covered, where the retained node is behaviorally indistinguishable.
+- **Error handling.** The `<packages>` guard is retained verbatim and now pinned by a dedicated test. No silent catches introduced.
+
+## Test Quality Assessment
+
+- **Fail-before discipline.** F1-F4 assert counts (`lines-valid`, `lines-covered`, `branches-valid`, `branches-covered`) precisely because the rates do not discriminate — the fail-before artifact records that F1's rate assertion passed against the defective code, validating the fixture-design-trap warning. Reviewer reproduced the four failures against the extracted base file with the exact recorded wrong values.
+- **Purity.** All 11 new `It` blocks use inline single-quoted here-strings cast to `[xml]`; zero disk writes, zero mocks in new tests, zero external dependencies. Grep for file-system APIs over the test file: no matches (the one `Mock` is a pre-existing allowlist test, untouched).
+- **Coverage of the new helper.** 39/40 lines (97.50%); all three precedence branches and the neither-element boundary have dedicated tests; the single gap is the Minor finding above.
+- **Non-regression.** Test-file diff is +246/-0; all 8 pre-existing `It` blocks pass unmodified (19/19 at HEAD, reviewer-run).
+- **F5/F6 as designed.** Both pass before and after the fix (guard fixtures, correctly not tagged expect-fail): F5 pins that method-only lines are not dropped by the union approach; F6 pins that `<methods>` subtrees are neither merged nor stripped, which is the recorded non-goal now filed as issue #530.
+
+## Verdict
+
+**PASS — 0 blocking findings; 2 Minor and 3 Informational non-blocking findings.** The change is merge-ready as reviewed; NF-1's recommended fixture can ride any later change to this file (or #530/#529 work) without holding this merge.

docs/features/active/2026-08-10-cobertura-coverage-arithmetic-441/evidence/baseline/assumption2-subset-proof.2026-08-10T22-30.md

@@ -0,0 +1,90 @@
+# Assumption 2 Subset Proof (P0-T18)
+
+Timestamp: 2026-08-10T22-30
+
+Proves `spec.md` § Assumptions item 2 — *method-level line numbers are a subset of the class-level
+rollup* — exhaustively on both committed sample documents, **before** any implementation change, so
+that document drift fails at baseline rather than after Phase 2.
+
+`spec.md` records this assumption as verified on three spot-checked classes only and explicitly
+`UNVERIFIED` across all classes, because exhaustive proof requires script execution. This artifact
+supplies that execution.
+
+Method: a **read-only streaming `XmlReader` pass** (not an `[xml]` cast — the inputs are 17.5 MB and
+10.4 MB) that applies, per `<class>`, exactly the union and `max(hits)` rule this plan specifies:
+enumerate both the class-level `./lines/line` axis and the `./methods/method/lines/line` axis, key
+by `[int]number`, resolve repeats by `max(hits)` / `branch=True` if either / `condition-coverage`
+from the larger `Total` tie-broken by larger `Covered`. A **method-only line key** is a key observed
+on the method axis and never on the class-level axis.
+
+Command:
+
+```powershell
+$root = (git rev-parse --show-toplevel) -replace '/', '\'
+& <scratchpad>\Test-Assumption2.ps1 -Path (Join-Path $root 'docs\features\active\2026-08-06-quickfiler-high-confidence-queue-init-stall-424\evidence\baseline\coverage-baseline.cobertura.xml')
+& <scratchpad>\Test-Assumption2.ps1 -Path (Join-Path $root 'docs\features\active\2026-08-06-quickfiler-high-confidence-queue-init-stall-424\evidence\qa-gates\coverage-final.cobertura.xml')
+```
+
+The analysis script is a throwaway agent-session script held outside the repository (scratchpad); it
+writes nothing, mutates nothing, and is not part of the change. Its full logic is described above.
+
+EXIT_CODE: 0
+
+Output Summary:
+
+```
+FILE=coverage-baseline.cobertura.xml
+classes=3169
+class-level distinct=79957
+union distinct=79957
+union covered=56124
+union branches valid=23109
+union branches covered=13472
+method-only line keys=0
+
+FILE=coverage-final.cobertura.xml
+classes=534
+class-level distinct=62345
+union distinct=62345
+union covered=53013
+union branches valid=15828
+union branches covered=12445
+method-only line keys=0
+```
+
+## `coverage-baseline.cobertura.xml` (raw generator output, 3169 classes)
+
+| Quantity | Required by P0-T18 | Measured | Match |
+| --- | --- | --- | --- |
+| class-level distinct | 79957 | **79957** | yes |
+| union distinct | 79957 | **79957** | yes |
+| union covered | 56124 | **56124** | yes |
+| union branches valid | 23109 | **23109** | yes |
+| union branches covered | 13472 | **13472** | yes |
+| **method-only line keys** | **0** | **0** | yes |
+
+## `coverage-final.cobertura.xml` (post-processed, 534 classes)
+
+| Quantity | Required by P0-T18 | Measured | Match |
+| --- | --- | --- | --- |
+| class-level distinct | 62345 | **62345** | yes |
+| union distinct | 62345 | **62345** | yes |
+| union covered | 53013 | **53013** | yes |
+| **method-only line keys** | **0** | **0** | yes |
+
+Additionally measured (not required by P0-T18, recorded as an observation): union branches valid
+**15828** and union branches covered **12445**. This falls inside the range `[15730, 16582]` that the
+research derived analytically for class-level `branches-valid` on this document, against the emitted
+defective value of 27848. It is an independent forward prediction of the post-change branch figures
+for this input.
+
+## Verdict
+
+**Assumption 2 holds exhaustively on both documents.** `union distinct == class-level distinct` and
+`method-only line keys == 0` on every one of the 3169 + 534 = 3703 classes, so the union design and
+the class-level oracle agree everywhere. The union formulation therefore reproduces the generator's
+own arithmetic exactly (79957 / 56124 / 23109 / 13472), which is the primary correctness oracle for
+AC-1.
+
+No spec-level finding is raised. No plan or spec revision is required. The Phase 5
+return-to-Phase-2 loop is not implicated by this task in any case.

docs/features/active/2026-08-10-cobertura-coverage-arithmetic-441/evidence/baseline/git-baseline.2026-08-10T22-30.md

@@ -0,0 +1,44 @@
+# Git Baseline — 2026-08-10-cobertura-coverage-arithmetic-441
+
+Timestamp: 2026-08-10T22-30
+
+Command:
+
+```powershell
+$root = (git rev-parse --show-toplevel) -replace '/', '\'
+Set-Location $root
+git rev-parse --abbrev-ref HEAD
+git rev-parse HEAD
+git rev-parse edf3d34c
+git merge-base --is-ancestor edf3d34c HEAD; 'ancestor-check-exit=' + $LASTEXITCODE
+git status --porcelain
+```
+
+EXIT_CODE: 0
+
+Output Summary:
+
+```
+bug/cobertura-coverage-arithmetic-441
+a5e336e5ae3443d4197caf5f87036fae1d538f89
+edf3d34cb9cd455bd3c1d9f5ee363b825632073c
+ancestor-check-exit=0
+--- porcelain ---
+ M docs/features/active/2026-08-10-cobertura-coverage-arithmetic-441/plan.2026-08-10T14-07.md
+?? docs/features/active/2026-08-10-cobertura-coverage-arithmetic-441/evidence/
+--- end porcelain ---
+```
+
+## Assessment
+
+- **Branch:** `bug/cobertura-coverage-arithmetic-441` — matches the plan's required branch exactly.
+  No branch is created, switched, or renamed by this plan.
+- **HEAD:** `a5e336e5ae3443d4197caf5f87036fae1d538f89`. Recorded as an observation only; it is never
+  used as a later expectation.
+- **Base commit:** `edf3d34c` resolves to `edf3d34cb9cd455bd3c1d9f5ee363b825632073c`.
+  `git merge-base --is-ancestor edf3d34c HEAD` exited **0**, confirming the base is an ancestor of
+  HEAD, so `git diff ... edf3d34c` gates are well-defined.
+- **Working-tree state at baseline:** the plan file
+  `plan.2026-08-10T14-07.md` is already ` M` (line-ending normalization performed by
+  `epic-orchestrator` before delegation) and the feature `evidence/` directory is untracked because
+  this run created it. Both source files under `scripts/` and `tests/` are clean at baseline.