chore(memory): consolidate remaining stranded worktree content from cleanup pass

.claude/agent-memory/atomic-executor/MEMORY.md

@@ -71,6 +71,7 @@
 - [pwsh -Command quoting + backtick stripping](project_pwsh_command_quoting_from_bash.md) — backticks vanish; `-split "\r\n"` gives 1 element
 - [pwsh -File binds a list as ONE string](project_pwsh_file_array_param_from_bash.md) — `-Tokens a,b,c` is a 1-element array
 - [Compile-time red needs body-level refs](project_compile_red_needs_body_level_references.md) — a missing type in a SIGNATURE hides body errors
+- [Cross-task shell-variable splat gates](project_cross_task_shell_variable_splat_gate.md) — a bare `@assemblies` in a fresh session runs zero assemblies and still reports zero failures; require same-payload enumeration + a `>= 1` count floor
 - [Shared evidence artifact + floating <ts>](project_shared_evidence_artifact_floating_ts.md) — N tasks told to append to "the same" artifact whose
 ## Test execution & isolation
 - [Long runs need a detached process](project_long_runs_need_detached_process.md) — background runners die at ~1h

.claude/agent-memory/atomic-executor/project_cross_task_shell_variable_splat_gate.md

@@ -0,0 +1,24 @@
+---
+name: cross-task-shell-variable-splat-gate
+description: A plan task that consumes a shell variable produced by an earlier task is vacuous unless the enumeration body is re-executed in the SAME -Command payload and a count-identity gate plus a >= 1 floor is asserted.
+metadata:
+  type: project
+---
+
+When a plan enumerates test assemblies in task A (`$assemblies = @(...)`) and splats them in task B
+(`& $vstest @assemblies ...`), no shell state persists between tool invocations, so a bare
+`@assemblies` in task B expands to nothing. vstest then runs zero assemblies and still reports a
+zero failure count, so the gate passes while measuring nothing.
+
+**Why:** #491 preflight round 2 blocked on exactly this. The fix that cleared round 3 was two-part:
+(1) task B must re-execute task A's enumeration BODY (not the whole `pwsh -NoProfile -Command '...'`
+wrapper — nesting pwsh puts the variable in a child session) inside the same payload; (2) acceptance
+must assert the assembly count on the command line equals the count recorded in task A's artifact
+AND is at least 1. The `>= 1` conjunct is what makes it falsifiable; the equality alone normally
+always holds because nothing between the two tasks adds a test project.
+
+**How to apply:** Whenever a plan passes a shell variable across task boundaries, demand both the
+same-payload clause and the `>= 1` floor. Also check the composed payload's quoting: a vswhere
+re-resolution embeds single quotes (`& 'C:\Program Files (x86)\...'`) and needs them doubled inside
+a single-quoted `-Command` payload. See [[project_pwsh_command_quoting_from_bash]] and
+[[project_preflight_selfderived_gate_thresholds_are_blind]].

.claude/agent-memory/atomic-planner/MEMORY.md

@@ -7,6 +7,7 @@
 - [#501 R3](project_501_r3_preflight_seams.md) — repo-wide 0-skipped gates unsatisfiable; BASELINE_FAILURE_SET subset; `Task.CompletedTask` singleton
 - [#511 R1](project_511_r1_preflight_delta_seams.md) — mid-cycle evidence deletion; git-log scans post-commit; Start-Process for 20-min runs
 - [#484](project_484_qfc_revision_seams.md) — ownership change sweeps plan→issue.md→spec.md (spec is the AC source); old-cardinal grep sweep
+- [QfcItemController test-capacity squeeze](project_qfcitemcontroller_test_capacity_squeeze.md) — four owned test files hold only ~471 aggregate spare lines and `.csproj` edits are barred; budget in Phase 0, permit relocation, mandate DataRow/shared-arrange compaction
 - [#494](project_494_threshold_reconciliation_plan_seams.md) — coverage runner throws before post-processing; reported-floor must not become hook-Blocking
 - [#498](conditional-ladder-and-unowned-class-gates.md) — gate EVERY rung of a recorded-selector ladder incl. rung 1; scope failing-identifier clauses to owned test classes; 0/0 changed-line figure → NOT APPLICABLE
 - [#503](project_503_ribbon_readiness_plan_seams.md) — RibbonViewer 487/500 forces a region move; 6+4 Compile entries; compile-time red + dossier
@@ -40,6 +41,7 @@
 - [Wiring gates must be wiring-sensitive](feedback_wiring_gates_must_be_wiring_sensitive.md) — count floors deflate with the defect they guard
 - [Research claims as acceptance clauses](research-claims-as-acceptance-clauses.md) — never encode an unmeasured third-party claim as a literal AC clause
 - [Literal-call clauses block file-size tightening](literal-call-clauses-block-file-size-tightening.md) — pinning a call in 2+ places near a 500-line file is unsatisfiable
+- [Enumeration variable must match its consumer](enumeration-variable-must-match-consumer.md) — `$kept` produced vs `@assemblies` splatted = zero-assembly vstest run reporting zero failures; same-payload re-execution + a count-parity floor
 - [Diff gates need a commit task](diff-gates-need-a-commit-task.md) — `git diff <BASE>..HEAD` passes vacuously with no commit task
 - [Never pin a HEAD SHA as a plan expectation](never-pin-head-sha-as-plan-expectation.md) — gate on tree invariants instead
 - [Absolute counts in shared files go stale](absolute-counts-in-shared-files-go-stale.md) — lower-bound/baseline-relative for co-owned files; keep exact the count the task changes
@@ -48,7 +50,7 @@
 
 ## C# toolchain and test mechanics
 
-- [Phase 0 toolchain bootstrap](project_csharp_phase0_toolchain_bootstrap.md) — global csharpier/dotnet-coverage exe + mandatory NuGet restore
+- [Phase 0 toolchain bootstrap](project_csharp_phase0_toolchain_bootstrap.md) — `dotnet tool run csharpier` works once the SDK is bootstrapped (global.json's missing .dotnet-sdk was the real blocker, not the manifest); mandatory NuGet restore
 - [Agent worktrees need SDK + NuGet + analyzer backfill](agent-worktrees-need-sdk-and-nuget-bootstrap.md) — four Phase 0 steps; CS0006 is an error, not a warning
 - [vstest scoped-run + csharpier 1.2.6 commands](reference_vstest_scoped_run_command.md) — vswhere + `/InIsolation` + `/TestCaseFilter`; csharpier needs a subcommand
 - [CSharpier gate: format not pipe-files](csharpier-format-not-pipe-files-gate.md) — `pipe-files` is stdout-only and non-enforcing
@@ -64,6 +66,8 @@
 - [Pester exits 0 on failing It blocks](pester-invoke-does-not-exit-nonzero.md) — scope every exit-code clause to a named channel
 - [PowerShell gate observables](powershell-gate-observables.md) — no Invoke-Pester exit code; explicit `scan_folders`; aggregate-only `CoveragePercent`
 - [Legacy csproj wiring](project_legacy_csproj_explicit_compile_include.md) — `Compile Include` + own `Reference`; ProjectReference gives no compile-time flow
+- [Invoke-VSBuild rewrites csproj HintPaths](invoke-vsbuild-rewrites-csproj-hintpaths.md) — the wrapper runs Sync-PackageReferences over EVERY csproj; with a forbidden .csproj the build itself commits the scope violation — use vswhere-resolved MSBuild
+- [Declaration-only seam task for fail-before](declaration-only-seam-task-for-fail-before.md) — tests citing not-yet-existing internals redden the whole assembly; order compile-clean tests first, open the fix phase with a no-behaviour seam task + whole-set assertion-time red run
 - [net48 / nullable context mismatch](project_nullable_context_mismatch_prod_vs_test.md) — check `#nullable enable` in prod AND missing `<LangVersion>` in the test csproj
 - [Worktree root breaks the `\.claude\` exclusion](worktree-root-breaks-dotclaude-exclusion.md) — assert a workspace-root prefix instead
 
@@ -80,7 +84,10 @@
 - [#511 R1 preflight delta seams](project_511_r1_preflight_delta_seams.md) — mid-cycle raw-evidence deletion breaks resolves gates; git-log scans post-commit only; absolute MSBuild path; Start-Process mechanic for 20-min runs; per-class coverage noise -0.50pp
 - [CSharpier "Formatted N files" is processed count](csharpier-formatted-n-is-processed-count.md) — restart-on-rewrite loops keyed on it never terminate; define rewritten-count via before/after SHA-256
 - [Terminal-phase planner traps](terminal-phase-planner-traps.md) — sweep the last phase for an unowned "a follow-up issue should carry it", artifacts written after the clean-tree commit task, and a false "clarification against the spec's wording"
+- [#493 UiThread dispatcher plan seams](project_493_uithread_dispatcher_plan_seams.md) — signature-change fail-before gets a REAL red build by staging the two `<Compile Include>` lines; the coverage script IS the parallelized run; scope `SwapUiThreadDispatcher`/`UiThreadDispatcherGate` greps to `QuickFiler.Test/`
 - [#553 CI parallel-split plan seams](project_553_ci_parallel_split_plan_seams.md) — workflow-only scope: no C# toolchain; Phase 0 snapshot for byte-identity; ruleset PUT + gh pr create orchestrator-gated; no jq (ConvertTo-Json -Depth 20); pathspec anchoring; BRANCH/SCRATCH conventions
+- [#442 QuickFiler metrics plan seams](project_442_quickfiler_metrics_plan_seams.md) — commented-out code defeats zero-hit grep gates; an AC conjunct already green pre-fix; declare the seam BEFORE the red tests or they don't compile; Invoke-MSTestWithCoverage throws below 80%
+- [#468 QfcCollectionController plan seams](project_468_qfc_collection_controller_plan_seams.md) — ToggleUnGroupConv is not COM-free drivable; MakeSpaceForItems never touches Size; a sign-defect seam must land carrying the defect; `LoadItemGroup(` needs the paren
 - [Threshold conflict: CLAUDE.md vs general-unit-test.md](project_coverage_threshold_conflict_claude_md_vs_general_unit_test.md) — 80/90 vs 85/75; repo-wide figure non-blocking, change-scoped gates blocking
 - [JaCoCo hook, Cobertura also accepted](project_csharp_coverage_gate_jacoco_format.md) — follow the format the delta names
 - [Async state machines split the denominator](async-state-machine-coverage-aggregation.md) — aggregate by `filename` or a >=90% gate fails for measurement reasons

.claude/agent-memory/atomic-planner/agent-worktrees-need-sdk-and-nuget-bootstrap.md

@@ -15,7 +15,11 @@ before the first `dotnet tool restore` and before the first `msbuild`, in this o
    `pwsh -NoProfile -File .\scripts\vscode\Install-RepoDotNetSdk.ps1`, or mirror the populated
    `.dotnet-sdk` tree from the main checkout. Falsifiable acceptance: `dotnet --version` prints
    `8.0.205` AND `dotnet --list-sdks` includes a path ending `.dotnet-sdk\sdk`.
-2. **`nuget restore TaskMaster.sln`.** A fresh worktree has no `packages/`. Every project declares
+2. **`nuget restore TaskMaster.sln`.** Prefer `pwsh -NoProfile -File .\scripts\vscode\Invoke-Restore.ps1`,
+   which resolves MSBuild through `vswhere` and runs `/t:Restore /p:RestorePackagesConfig=true /m` — the
+   `packages.config`-aware form. `nuget.exe` is not guaranteed to be on `PATH` in an agent worktree, so a
+   plan task whose only stated command is `nuget restore` can be unrunnable.
+   A fresh worktree has no `packages/`. Every project declares
    `<Target Name="EnsureNuGetPackageBuildImports" BeforeTargets="PrepareForBuild">` whose `<Error>`
    fires before compilation when the tree is missing (e.g. `QuickFiler.Test.csproj:452-466`), and every
    `Reference` `HintPath` under `..\packages\` is unresolvable. CI does not hit this because

.claude/agent-memory/atomic-planner/declaration-only-seam-task-for-fail-before.md

@@ -0,0 +1,19 @@
+---
+name: declaration-only-seam-task-for-fail-before
+description: When regression tests reference internal seams that do not exist yet, the whole test ASSEMBLY goes red and no test in it can run — order compile-clean tests first, then open Phase 2 with a declaration-only seam task so every test fails at assertion time
+metadata:
+  type: feedback
+---
+
+A C# `[expect-fail]` phase whose tests reference not-yet-existing `internal` members produces a compile error, not a failing test. Because the failure is per-ASSEMBLY, it also destroys the assertion-time evidence of every other test in that phase, including ones that would have compiled fine.
+
+Two ordering rules fix it:
+
+1. **Order the fail-first phase so every compile-clean test is authored and RUN before the first compile-breaking one.** Reflection/structural tests and argument-guard tests against existing members compile today; marshalling tests through a new internal constructor and state assertions through a new internal observation member do not. Run and record the former group first.
+2. **Open the fix phase with a declaration-only seam task.** It declares exactly the new members (constructor overload, `internal bool` observation properties, nullable collaborator field) with no behaviour change, and its acceptance is `EXIT_CODE: 0` on a `/t:Build`. The next task then runs the WHOLE test set and records "N discovered, N failed, 0 passed, 0 build errors" — that artifact, not the compile errors, is the authoritative fail-before record, and no behavioural fix precedes it.
+
+Give each compile-time-red task its own falsifiable acceptance anyway: "a `[TestMethod]` named `<X>` exists in `<file>` AND the build is red with the error list recorded verbatim". The method name is a distinct literal per task, so two tasks that share the same broken build are still separately checkable. Quote every such method name in a plan prose block so rule G5 exonerates it.
+
+**Why:** on the #476 plan, six of eight Phase 1 tasks referenced `internal WebView2BreadcrumbHost(` / `IsAttached` / `HasUiDispatcher`. Without this shape the plan would have claimed a fail-before it could not show, and a preflight reviewer cannot tell a genuine assertion failure from a typo that broke the build.
+
+**How to apply:** whenever a bug plan's tests need a NEW seam, sort the phase by "compiles against HEAD" first, then add the declaration-only task and the whole-set fail-before run as the first two tasks of the fix phase. Related: [[acceptance-edits-must-be-false-before-true-after]], [[one-ac-per-checkoff-task]].

.claude/agent-memory/atomic-planner/diff-gates-need-a-commit-task.md

@@ -15,4 +15,6 @@ Any plan gate expressed as `git diff --numstat <MERGE_BASE>..HEAD` or `git diff
 2. **Phase 0 porcelain is non-empty by construction.** Never write "empty `git status --porcelain`" as a Phase 0 acceptance. Record the verbatim output and make the binary outcome a *type* assertion instead: no `.cs`, `.csproj`, `.xml`, or `.sln` path appears. Docs/evidence/agent-memory paths are expected.
 3. **Scope-lock diff gates must whitelist docs.** Once commits exist, the diff legitimately contains `docs/features/` and `.claude/agent-memory/` paths. Phrase the gate as "every path is a scope-lock member OR under docs/evidence" and keep the hard clause on source extensions only. In a companion file-size audit, state that Markdown docs are exempt from the 500-line cap per `.claude/rules/general-code-change.md`.
 
+4. **Branch history can make a `base..HEAD` gate unsatisfiable no matter how it is whitelisted.** An epic child branch often already carries commits the child does not own — #491's branch head sat on a merge of `main` that restored 58 `docs/features/potential/` records, so a P5 scope-lock gate enumerating five `QuickFiler.Test/**` paths failed before the executor touched anything. When the gate's intent is "this child changed only its own paths", gate the child's OWN commit with `git show --name-only --format= HEAD` run right after the plan's commit task, and demote the branch-wide `base..HEAD` list to a recorded observation rather than a pass/fail condition.
+
 Related: [[never-pin-head-sha-as-plan-expectation]] — the fix is commit tasks plus tree-invariant gates, never a pinned SHA equality check.

.claude/agent-memory/atomic-planner/enumeration-variable-must-match-consumer.md

@@ -0,0 +1,17 @@
+---
+name: enumeration-variable-must-match-consumer
+description: A plan task that enumerates test assemblies must assign the exact variable name later vstest tasks splat, and enumeration plus invocation must sit in one pwsh -Command payload — otherwise vstest runs on zero assemblies and still reports zero failures
+metadata:
+  type: feedback
+---
+
+When one task computes a list and a later task splats it (`& $vstest @assemblies ...`), the producing command must assign that exact name, and the producer's body must be re-executed inside the SAME `pwsh -NoProfile -Command` payload that invokes the consumer.
+
+**Why:** #491 preflight revision 3. P0-T17 assigned `$all`, `$claude`, and `$kept`; P0-T18 and P3-T6 both splatted `@assemblies`. An undefined `@assemblies` expands to nothing, so vstest would have run against ZERO assemblies and still reported a zero failure count — a silently green full-suite gate. The plan's own conventions paragraph already stated the no-shell-state-persists rule and still shipped the mismatch, because the rule was written about session lifetime and nobody re-read the producing command for the name it actually emits. P0-T17 is itself a `pwsh -Command` child process, so no variable it sets can survive into a later tool invocation regardless.
+
+**How to apply:** Two checks before a plan leaves preflight.
+
+1. **Name check.** Grep the plan for every `@<name>` splat and every `$<name>` consumption, then confirm the producing command literally assigns that identifier. Renaming the producer's variable is the cheaper fix than renaming every consumer; a display label such as `KEPT={2}` in a format string may stay as-is and should be called out as deliberate so a reviewer does not flag the mismatch.
+2. **Payload check.** Every consuming task's text must say explicitly that it re-executes the enumeration body verbatim in the same payload, and that a bare `@<name>` in a fresh session is forbidden. Add an acceptance clause pinning the count actually passed on the command line to the count recorded in the enumerating task's artifact, with a `>= 1` floor — otherwise the empty-splat failure mode has no gate that can catch it.
+
+Related: [[wiring-gates-must-be-wiring-sensitive]] — a count floor is the observable that makes an empty run fail; [[powershell-gate-observables]] for the quoting rules of the enclosing payload.

.claude/agent-memory/atomic-planner/invoke-vsbuild-rewrites-csproj-hintpaths.md

@@ -0,0 +1,17 @@
+---
+name: invoke-vsbuild-rewrites-csproj-hintpaths
+description: Invoke-VSBuild.ps1 runs Sync-PackageReferences.ps1, which rewrites <HintPath> in EVERY csproj — so when a feature has a forbidden .csproj, the build wrapper itself can commit the scope violation; use direct MSBuild via vswhere instead
+metadata:
+  type: feedback
+---
+
+When a plan carries a forbidden-file list that includes a `.csproj`, do NOT plan `scripts/vscode/Invoke-VSBuild.ps1`. Resolve `MSBuild.exe` through `vswhere` and pass the `CLAUDE.md` argument list directly.
+
+**Why:** `Invoke-VSBuild.ps1:152-155` unconditionally invokes `scripts/vscode/Sync-PackageReferences.ps1 -SolutionRoot $repoRoot` before building. That script scans every `*.csproj` for `<HintPath>..\packages\...\lib\...</HintPath>` values that do not resolve on disk and rewrites them in place (`Sync-PackageReferences.ps1:55-154`). In a fresh agent worktree — where `packages/` is absent or holds versions skewed from `packages.config` — that is a live rewrite path, not a theoretical one. On #476 the forbidden list included `QuickFiler/QuickFiler.csproj`, so a single wrapper-driven build would have produced a scope violation that the feature's own scope-containment gate then reports against the executor.
+
+Two related facts:
+
+- The wrapper otherwise supports everything the mandated commands need (`-Target Rebuild`, `-EnableNETAnalyzers`, `-EnforceCodeStyleInBuild`, `-TreatWarningsAsErrors`), so the reason to skip it is exclusively the csproj rewrite. The general "prefer repo-defined tasks" guidance in `policy-compliance-order` still holds everywhere else.
+- The wrapper `throw`s on a non-zero MSBuild exit (`:165-167`), so a `pwsh -File` run of it exits 1 rather than surfacing MSBuild's own code. `scripts/vscode/Invoke-MSTestWithCoverage.ps1:236` does the same. Baseline tasks must take `EXIT_CODE:` from the pwsh process and must not gate on `EXIT_CODE: 0`.
+
+**How to apply:** grep a draft plan for `Invoke-VSBuild`; if the feature forbids any `.csproj`, swap to the vswhere-resolved `& $msbuild TaskMaster.sln /t:Rebuild ...` form and say in the plan why. Also add a Phase 5 gate that no `.csproj` outside the writable set appears in the change inventory. Related: [[poshqc-mcp-and-msbuild-invocation-facts]], [[project-csharp-phase0-toolchain-bootstrap]].