docs: re-measure #218/#220/#222/#223/#224 delta, verify #225/#228 on real repos

docs/notes/precision-remeasure-2026-07-11.md

@@ -0,0 +1,215 @@
+# Precision re-measure — 2026-07-11 (verifying #218/#220/#222/#223/#224 and #225/#228)
+
+Follow-up to [`oracle-sweep-2026-07-10.md`](oracle-sweep-2026-07-10.md) (issue \#201),
+which drafted 8 precision-gap issues from real own-only findings in 5
+OSS repos: #218-#225. Four follow-up PRs have since shipped fixes:
+
+- [#230](https://github.com/PhysShell/Own.NET/pull/230) — Closes **#218**
+  (not #219 — #219 is the separate WinForms `Controls`/`IContainer` disposal
+  channel gap, still open). DependencyProperty/property-changed old→new
+  subscription-rotation recognition.
+- [#231](https://github.com/PhysShell/Own.NET/pull/231) — four smaller gaps
+  in one PR: `CommandManager.RequerySuggested` allowlist (#223),
+  `using (field = new T())` release (#220), template-part local/pattern-var
+  exemption (#222), self-detaching handler (#224).
+- [#232](https://github.com/PhysShell/Own.NET/pull/232) — Closes **#228**
+  (sub-part (b) of the parent #221 split), curated app-scoped source
+  exemption for an `Application`-derived subscriber.
+- [#233](https://github.com/PhysShell/Own.NET/pull/233) — Closes **#225**,
+  exempt locals of user types whose `Dispose()` is provably empty in source.
+
+This note has two parts: **Part 1** formalizes yesterday's measurement of
+\#230/\#231 against the 2026-07-10 baseline (same method as before, now with
+the #218/#219 numbering corrected). **Part 2** reruns the same method against
+today's fresh `main` (with #232/#233 also merged) for the two repos those
+fixes target, and checks the actual delta against the expected delta each
+PR's own "validation plan" section predicted.
+
+## Method (unchanged from the 2026-07-10 rerun)
+
+- Same repo clones, same pinned commits as the original sweep: MahApps.Metro
+  `72099e3`, MaterialDesignInXamlToolkit `ef3a5ea`, AvalonEdit `ed0bd14`,
+  ShareX `0df9ca4`, ClosedXML `4e89dce` — no drift from upstream, an
+  apples-to-apples comparison.
+- `OwnSharp.Extractor` built fresh off the `main` commit under test.
+- `scripts/own-check.sh --format sarif --severity warning -- <repo>`, same
+  `OWN_EXTRA_REF_DIRS` (WindowsDesktop ref pack) as `oracle.yml` uses.
+- SARIF results diffed by normalized `(file, line)` per rule ID between the
+  "before" and "after" runs.
+
+## Part 1 — #218/#220/#222/#223/#224 delta (`main` at `3ca4bde`, #230+#231 merged)
+
+| repo | OWN001+OWN014 before | after | Δ | new findings introduced |
+|---|---:|---:|---:|---:|
+| MahApps.Metro | 3 | 1 | **−2** | 0 |
+| MaterialDesignInXamlToolkit | 23 | 19 | **−4** | 0 |
+| AvalonEdit | 24 (23 OWN001 + 1 OWN014) | 17 | **−7** | 0 |
+| ShareX | 235 | 232 | **−3** | 0 |
+| **total** | **285** | **269** | **−16** | **0** |
+
+Zero new findings anywhere. OWN050 (unresolved-reference advisory) counts are
+byte-for-byte identical before/after in all 4 repos — coverage didn't shift,
+only the fixes' targets did.
+
+### Removed findings, mapped to the fix that closed them
+
+| repo | site | issue / PR | rule |
+|---|---|---|---|
+| MahApps.Metro | `Actions/CommandTriggerAction.cs:116` | **#218** / #230 (DP rotation) | OWN001 |
+| MahApps.Metro | `Controls/MetroWindow.cs:1448` | #222 / #231 (template-part pattern var) | OWN001 |
+| MaterialDesignInXamlToolkit | `SmartHint.cs:205-208` (×4) | **#218** / #230 (DP rotation) | OWN001 |
+| AvalonEdit | `Editing/AbstractMargin.cs:99` | **#218** / #230 (DP rotation) | OWN001 |
+| AvalonEdit | `Editing/LineNumberMargin.cs:114` | **#218** / #230 (DP rotation) | OWN001 |
+| AvalonEdit | `Folding/FoldingMargin.cs:218` | **#218** / #230 (DP rotation) | OWN001 |
+| AvalonEdit | `CodeCompletion/OverloadViewer.cs:58,64` | #222 / #231 (template-part local) | OWN001 |
+| AvalonEdit | `Search/DropDownButton.cs:78` | #224 / #231 (self-detaching handler) | OWN001 |
+| AvalonEdit | `Editing/ImeSupport.cs:47` (`CommandManager.RequerySuggested`) | #223 / #231 (weak-event allowlist) | **OWN014** |
+| ShareX | `Cryptographic/HashChecker.cs:42` | #220 / #231 (`using (field = new T())`) | OWN001 |
+| ShareX | `TaskEx.cs:41` | #220 / #231 | OWN001 |
+| ShareX | `IndexerJson.cs:35` | #220 / #231 | OWN001 |
+
+By fix: rotation (**#218**/#230) accounts for 8 of the 16 removed findings —
+beating the "−6+" estimate from the original sweep note (which only
+individually verified 6 sites; the rerun confirms all 8 rotation sites in
+the baseline actually clear). The other three #231 fixes account for the
+remaining 8: template-part locals (3), self-detach (1), weak-event allowlist
+(1, on the OWN014 hard-error tier as designed), using-field release (3).
+
+### Real bugs — confirmed still flagged
+
+ShareX `HistoryLib/HistoryItemManager_ContextMenu.cs` cluster (48 findings —
+no `Dispose()` at all), ShareX `ShapeManagerMenu.cs` cluster (the flagship
+`menuForm` leak), ShareX plain `Timer`/`ImageList` fields
+(`ColorUserControl.cs`, `AutoCaptureForm.cs`, `UploadersConfigForm.cs`), and
+AvalonEdit `TextView.cs:1843,1946` — all unchanged, still flagged.
+
+### Still open (out of #230/#231's scope)
+
+MahApps.Metro `Behaviors/TiltBehavior.cs:70` (self-owned `Behavior<T>`
+source, #227) and MaterialDesignInXamlToolkit's remaining 19 findings —
+`App.xaml.cs:22`, `ListsAndGridsViewModel.cs` (#229), `CircleWipe.cs`/
+`FadeWipe.cs`, the 7 bulk-review `themeManager.ThemeChanged` sites, etc. —
+untouched, as expected.
+
+## Expected delta of the next measurement (before rerunning)
+
+Predicted from #232's and #233's own PR bodies, before this note's Part 2
+below actually reran anything:
+
+- **After #232 (Closes #228):** MaterialDesignInXamlToolkit
+  `MahMaterialDragablzMashUp/App.xaml.cs:10,22` (`themeManager.ThemeChanged`
+  on the `App`-derived subscriber) should drop out of own-only.
+- **After #233 (Closes #225):** ClosedXML `Excel/Cells/Slice.cs:91,109,149,188,219`
+  — five `Slice.Enumerator`/`ReverseEnumerator` locals, all sharing the
+  same empty-`Dispose()` root — should drop out of own-only.
+- **Not yet fixed (#227, #229 — still open):** MahApps.Metro
+  `Behaviors/TiltBehavior.cs:70` and MaterialDesignInXamlToolkit
+  `ListsAndGridsViewModel.cs:16-17` (both demo variants) should remain
+  flagged.
+
+## Part 2 — actual #232/#233 delta (`main` at `b1ee961`, #232+#233 also merged)
+
+| repo | before (after #230/#231) | after (+ #232/#233) | Δ | expected Δ |
+|---|---:|---:|---:|---:|
+| MaterialDesignInXamlToolkit | 19 | 18 | **−1** | −1 (1 site) |
+| ClosedXML | 270 | 5 | **−265** | −5 (1 site's 5 locals) |
+
+### MaterialDesignInXamlToolkit — matches expectation
+
+`MahMaterialDragablzMashUp/App.xaml.cs:22` (the only SARIF-reportable line
+for that site — line 10 in the issue's citation is the `PaletteHelper()`
+resolver-call context line, not a second finding) dropped out, exactly as
+predicted. No other site moved. **Match.**
+
+### ClosedXML — does NOT match expectation, and not in the way "more FPs cleared than expected" sounds
+
+The raw delta (−265) looks like a huge overachievement of the predicted −5,
+but breaking it down site-by-site shows it is **not** a clean superset —
+part of it is a real coverage gap, and the bulk of it is a genuine, unsound
+over-exemption that needs its own follow-up issue. Do not read "−265" as
+"the fix worked great"; read the breakdown below.
+
+**1) Only 2 of the 5 predicted Slice.cs sites actually cleared:**
+
+| line | local | enumerator type | `Dispose()` form | cleared? |
+|---:|---|---|---|:---:|
+| 91 | `enumerator` | `Enumerator` | `void IDisposable.Dispose() { }` — **explicit interface implementation** | ❌ still flagged |
+| 109 | `cellEnumerator` | `Enumerator` | same explicit-interface form | ❌ still flagged |
+| 149 | `cellEnumerator` | `Enumerator` | same explicit-interface form | ❌ still flagged |
+| 188 | `cellEnumerator` | `ReverseEnumerator` | `public void Dispose() { }` — plain method | ✅ cleared |
+| 219 | `enumerator` | `ReverseEnumerator` | same plain-method form | ✅ cleared |
+
+`HasEmptyDisposeBody` (the check #233 shipped) recognizes a plain
+`public void Dispose() { }` declaration but not an explicit interface
+implementation (`void IDisposable.Dispose() { }`) — same empty body,
+different declaration syntax. This is a **coverage gap** (the check is
+too narrow, not unsound): `Enumerator` and `ReverseEnumerator` sit right
+next to each other in the same file with the identical empty body, and
+only one form is recognized.
+
+**2) The other 263 removed findings are an unrelated, unsound over-exemption:**
+every one of them is a `wb`/`workbook`/`wb1`/`wb2`/`wb_saved`/`wbSource`
+local of type `XLWorkbook`, scattered across `ClosedXML.Examples` (263
+call sites in total) — nothing to do with `Slice.cs` at all. The root
+cause: `ClosedXML/Excel/XLWorkbook.cs:874`
+
+```csharp
+public void Dispose()
+{
+    // Leave this empty so that Janitor.Fody can do its work
+}
+```
+
+is empty **in source only**. `Janitor.Fody` is a build-time IL weaver that
+rewrites this exact method body at compile time to call the adjacent
+`DisposeManaged()` (`XLWorkbook.cs:868`, which does real cleanup —
+`Worksheets.ForEach(w => (w as XLWorksheet).Cleanup())`). A Roslyn-source
+analyzer has no visibility into post-compile IL weaving, so
+`HasEmptyDisposeBody`'s "zero-statement block ⇒ no-op" heuristic reads the
+comment-explained placeholder as a genuine no-op and exempts every
+`XLWorkbook` local in the compilation — even though disposing one
+(pre-weaving semantics aside) is exactly what real ClosedXML code is
+supposed to do, and never calling it is a real, if fairly benign
+(process-exit-scoped example code), leak.
+
+This is **not** what #233 set out to do (#233's own PR body scoped the fix
+to "only locals," and its acceptance checklist's regression diff was run
+against the existing sample corpus, which has no Fody-woven-Dispose type in
+it — so this shape never showed up in that PR's own local verification).
+It's a real precision regression surfaced only by running against actual
+OSS code that uses a code-weaving library, exactly the kind of gap a
+synthetic-sample-only regression suite can't catch.
+
+**Confirmed unaffected (real findings, unrelated to this fix):**
+`ClosedXML.Tests/Excel/CalcEngine/FunctionsTests.cs:285` (`cts`, a genuine
+undisposed `CancellationTokenSource`) and
+`ClosedXML.Tests/Excel/Ranges/UsedAndUnusedCellsTests.cs:10` (`workbook`
+field, real) — both still flagged, so the over-exemption is scoped to
+`XLWorkbook` **locals** specifically (matches #233's own "only locals, not
+fields" boundary), not a wholesale collapse of ClosedXML's findings.
+
+### Net read for the next iteration (analyzer code NOT touched in this note)
+
+Two separate, differently-shaped follow-ups for whoever picks this up next:
+
+1. **Coverage gap** — extend `HasEmptyDisposeBody` to also recognize an
+   empty **explicit interface implementation** `Dispose()`
+   (`void IDisposable.Dispose() { }`), not just a plain public method
+   declaration. Closes the remaining 3/5 Slice.cs sites (91, 109, 149).
+2. **Soundness regression** — `HasEmptyDisposeBody` needs to stop treating
+   a literally-empty `Dispose()` as proof of "nothing
+   to release" when IL-weaving tools (Janitor.Fody being the concrete
+   instance found here; PostSharp and other Fody plugins follow the same
+   shape) are in play. A cheap, conservative signal: if the type or any
+   partial-class fragment also declares a private method whose body the
+   weaver targets by convention (Janitor.Fody's own convention is a method
+   named `Dispose(Managed|Unmanaged)`/`DisposeManaged`/`DisposeUnmanaged`
+   living alongside the empty `Dispose()`), don't exempt — treat it as
+   "provably non-empty at runtime" and fall back to today's flagged
+   behavior. Whether that specific heuristic is the right one is a design
+   call for whoever picks up the fix, not decided here.
+
+This note's own instructions were read-only with respect to the analyzer —
+no code was changed to investigate or produce the breakdown above; the
+breakdown is entirely from comparing old/new SARIF output and reading the
+target repos' real source.