fix(635): settle the residual reflective-caller risk with a measured, non-vacuous audit

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

@@ -17,6 +17,7 @@
 - [#614](project_614_store_root_leak_plan_seams.md) — AC25 net non-growth; behavior-preserving seam phase reconciles fail-before with a signature change; net48 `IsNullOrWhiteSpace` doesn't narrow (`archiveRoot!`)
 - [#464 R3/R4](project_464_efc_controller_plan_seams.md) — additive-only file grows, budget a ceiling not a shrink; non-comment literal counts; a phase-N count must survive phases 1..N-1 deletions
 - [#677 R1–R8](project_677_keyboard_focus_leak_plan_seams.md) — ctor param REJECTED (5 reflection-arity tests); typed harness for compile-red; internal 9-arg ctor, never ambient SetSynchronizationContext; per-file non-vacuity floors
+- [#635](project_635_reflective_caller_audit_plan_seams.md) — evidence-only audit: tracked plan inflates its own sweep; scan hits its own pattern list; spec said six sites, tree has eight; pathspec breadth inflates a count
 - [#680](project_680_menu_mode_plan_seams.md) — HostTests.cs 499 not 500; set-difference format gate; TRX 5-shape identifiers, `grep -a`; append-a-dated-literal discriminator; post-merge remediation: exact line arithmetic — the review's "optional" fallback was load-bearing (501 vs 500)
 
 ## Plan-structure traps

.claude/agent-memory/atomic-planner/project_635_reflective_caller_audit_plan_seams.md

@@ -0,0 +1,46 @@
+---
+name: project-635-reflective-caller-audit-plan-seams
+description: Issue #635 evidence-only audit plan — self-scanning gates, tracked-plan-inflates-its-own-sweep, spec-vs-tree site-count mismatch, and pathspec breadth changing a reported count
+metadata:
+  type: project
+---
+
+Planning seams found while authoring the issue #635 residual-reflective-caller audit plan (a
+Markdown-only, evidence-producing audit that modifies no `.cs` file).
+
+**Why:** four defects were only visible by re-deriving the delegation's figures against the tree; each
+would have failed at execution.
+
+**How to apply:** on any plan whose deliverable is a repository-wide search recorded as evidence.
+
+1. **A tracked plan file inflates its own sweep.** The plan is tracked under the docs tree and quotes
+   every identifier it searches for, so a `git grep` partition that includes the docs tree returns
+   more hits after the plan is written than the base measurement the caller supplied. Never assert
+   that partition's total; assert the classification identity (per-category counts sum to the printed
+   total, the "genuine caller" category is `0`) and record the reason the total moved.
+
+2. **A host-identity / forbidden-token scan hits its own pattern list.** Both the scan's artifact and
+   the plan file quote the patterns verbatim, so a zero-hit gate over the feature folder is
+   unsatisfiable. Carve out exactly those two files by name filter, state the reason, and verify by
+   hand at planning time that they carry no real leak.
+
+3. **Enumerate from disk, not from the index, for a late-phase folder scan.** `git ls-files` misses
+   the Phase 4 artifacts that are still untracked; `Get-ChildItem -Recurse -File -Name` covers both
+   and prints folder-relative names, so no resolved provider path (which carries the account name)
+   reaches the artifact. `Select-String`'s `MatchInfo.Path` is resolved and absolute — print the
+   enumeration variable instead.
+
+4. **A spec's site count can disagree with the tree.** Spec AC-9 named "six variable-argument
+   reflection call sites"; the tree has eight against `typeof(QfcCollectionController)` (five
+   `GetField` plus one `GetMethod` in the test-support partial class, plus one in the navigation-digits
+   test file and one in the main test file). Enumerate the mechanically derived superset, note that it
+   names each of the spec's six individually and therefore discharges the AC, and record the
+   discrepancy as an evidence note rather than editing an approved spec.
+
+5. **A git pathspec is broader than a language.** `-- 'QuickFiler/*'` reaches tracked `.csproj`,
+   `.csproj.bak` and `packages.config`, so a `System.Reflection` count taken that way exceeds the
+   first-party `.cs` count (32) that a delegation may have measured. Assert a floor plus a total
+   classification whose classes include a manifest-reference class, never the bare number.
+
+Related: [[zero-hit-grep-gates-need-carveouts]], [[absolute-counts-in-shared-files-go-stale]],
+[[acceptance-edits-must-be-false-before-true-after]].

.claude/agent-memory/feature-review/MEMORY.md

@@ -78,6 +78,7 @@
 - [489-review-residuals](project_489-review-residuals.md) — cycle-1 reaudit GO/0 blocking; RC-1 cured (RED TRX committed with test, fix next commit = provable RED-first); commit-per-phase vs single-commit plan text judged on intent; amend the referenced table when an AC delegates to it
 - [493-review-residuals + msbuild-log gate adjudication](project_493-review-residuals-and-msbuild-log-gate-adjudication.md) — PASS/0 blocking; how a structurally unsatisfiable msbuild-log byte-equality gate was adjudicated without downgrading the AC
 - [680-review-residuals](project_680-review-residuals.md) — closed GO c3; leak class recurred 3x (c2 committed TRX, plan draft, pass-2 QA's OWN fresh vstest output) — fresh vstest TRX is born unsanitized, require in-task sanitize; re-run host sweep every cycle
+- [635-review-residuals](project_635-review-residuals.md) — GO/0 blocking; Markdown-only evidence audit; drift-invariant "total classification with one empty category" validated at a 3rd commit (2229->2337->2474, identities held); hook payload key is `output`; session-cwd mirror needed again
 - [677-review-residuals](project_677-review-residuals.md) — PASS/0 blocking; compile-red RED-first equivalence; 70.7% modified-file dispositioned non-blocking (below the #230 bar); owed: analyzer HintPath-skew + SetupDisposal-debt promotions; post-677 baseline 0.852804/0.792300
 
 ## Artifact hygiene

.claude/agent-memory/feature-review/project_635-review-residuals.md

@@ -0,0 +1,50 @@
+---
+name: 635-review-residuals
+description: Issue #635 reflective-caller audit review — GO/0 blocking; how a Markdown-only evidence-audit branch was verified, and the drift-invariant acceptance-condition pattern worth reusing
+metadata:
+  type: project
+---
+
+Issue #635 (`bug/issue-468-residual-reflective-caller-risk-635`, head `73bd8082`) reviewed
+2026-08-29: **GO, 0 blocking, 15/15 AC PASS**. Markdown-only diff, 32 paths, `NON_MD_COUNT: 0`.
+
+**Why:** the item is an evidence-producing audit discharging issue #468 AC-16. Its claims are negative
+results, so the review risk was an acceptance condition that could not have failed — not a code defect.
+
+**How to apply:**
+
+- **The drift-invariant acceptance-condition pattern is validated and worth reusing.** Partition B's
+  hit total moved 2229 (spec base) -> 2337 (execution) -> **2474 (my review head)** purely from prose
+  accretion, while both acceptance identities (`CAT_D + CAT_E = TOTAL`, `CAT_G = 0`) held at all three
+  commits. Expressing the condition as a total classification with one empty category, never a hit
+  count, is what made the evidence survive. A count-based condition would have been red twice over.
+- **Asserted vs non-asserted figures is the right split.** `SCOPE_FILES 683`, `AC16_SIX_EXTENSION_SCOPE
+  153`, `TRACKED_CS 1599`, Partition C `31` all reproduced exactly at review head, because the
+  Partition A pathspec excludes `docs/*` and `.claude/*` — the only two trees this branch writes into.
+  Drift in a deliberately non-asserted reference value is not a blocking finding.
+- **Verify the non-vacuity control actually discriminates.** P1-T2 ran the identical pathspec for a
+  present token and returned 13 hits / 4 files, two of them (`QuickFiler/Notes/notes_interface_hierarchy`,
+  extensionless; `QuickFiler.csproj.bak`) unreachable by any extension-based search. That is what makes
+  a zero a measurement rather than an artefact.
+- **Three claims had narrower supporting measurement than the claim; all held under the broader test I
+  ran.** (1) inventory used `Delegate.CreateDelegate`/`Activator.CreateInstance` vs the spec's bare
+  `CreateDelegate`/`Activator.` — bare forms also prod=0; (2) `dynamic` late binding never enumerated as
+  a mechanism, and it is the one class in the same family as the stated limit (runtime-assembled names)
+  — verified absent (1 prod comment hit, 0 in test tree); (3) assembly `ComVisible(false)` asserted to
+  bind every type — verified no per-type `ComVisible(true)`/`ClassInterface`/`ProgId`. Run the broader
+  form before accepting a narrowed pattern.
+- **AC-9 six-vs-eight:** spec names six variable-arg sites, mechanical derivation yields eight (7
+  `GetField(` + 1 `GetMethod(`), so no six-subset is identifiable. Recording (not editing the approved
+  spec, not silently picking a subset) was correct; AC discharged by superset. **Owed: maintainer
+  amendment of the spec baseline figure at merge.**
+- **The `//` ordering row is real:** `QuickFiler/Legacy/QuickFileController.cs:20` contains
+  `MethodBase.GetCurrentMethod()` but its first non-whitespace token is `//`, so ordered tests put it in
+  L3 not L1. Verified at character level.
+- **cwd hazard recurred** (see [[review-worktree-differs-from-session-cwd-mirror-artifacts]]): hook
+  exits 0 from the review worktree but **fails from the session cwd** `TaskMaster-wt/2026-08-29T00-11`,
+  which lacks the feature folder. Mirrored the 3 artifacts there; both cwds then exit 0. Hook payload
+  key is **`output`**, not `agent_output`.
+- **Coverage gate correctly not exercised:** no `artifacts/pr_context.summary.txt` -> `changedLanguages`
+  empty -> only the 3 path checks run. Zero changed files in every coverage language, so no coverage
+  artifact was emitted; emitting one would have fabricated a measurement and exposed an unrelated
+  repo-wide threshold.