test: re-derive the numbers and reasons four guards document - #2116

Closes #2110.

Four guards documented numbers or reasons that no longer matched the code they protect. None was failing, which is the point: a guard whose prose is wrong still passes, and the next reader takes its derivation as current fact.

1. SetCustom's slice ran ten times past its own body

EveryThemeEntryPoint_GoesThroughTheLegibilityCorrection bounded each method with "up to the next member declaration", implemented as find the next public, and only fall back to private if there is none. For a method followed by private members that skips over them to a later public one.

Measured: SetCustom's slice was 6727 characters against a real body of 675, swallowing IsDarkBackground, ApplyShade, Shade and Legible. Nothing in that region calls ApplyShade() today, so Assert.Contains("ApplyShade();", body) was still being satisfied by SetCustom's own call — but one added call anywhere in those six kilobytes and the guard could no longer tell whether SetCustom still does it.

Now brace-matched, via the BalancedBlock helper added in #2112.

2. Its justification was three fixes out of date

It said the assertion is on source rather than by calling the method "because SetCustom ends in Save(), which writes the user's real theme file". True before #1741 made SettingsPath redirectable — and false since: ThemeServiceTests calls SetCustom for real against a temp directory. The guard is still worth having, for a different reason (an entry point nobody wrote a test for), and now says so.

3. The admin-banner counts

Remarks and failure message both quoted "27 views", "27 of 27" and "29 of 29" — three numbers that disagreed with each other and with the tree. Re-derived: 62 banners across 31 views, 31 pairs, all on CornerRadius="12" Padding="12,8". The failure message now computes the count from what it parsed instead of quoting a remembered one.

4. The sidebar-label budget charged leaf rows for a deleted glyph column

The derivation subtracted "the 13px glyph and its 10px margin", leaving 155px. Leaf rows carry no glyph — NavItem has no such member, and the only two glyph bindings left in MainWindow.xaml are the group header's own and the single-item row reading its group's. A leaf label's real budget is 220 - 28 - 14 = 178px, about 27 characters. The asserted budgets (21 and 23) come from the longest shipping labels, not the pixel math, so they are unchanged.

5. The PR checklist omitted the one gate that fails after the merge

auto-release.yml requires the newest CHANGELOG heading to be dated today in UTC, and it runs after the squash merge, when the branch is gone. A stale date fails the release rather than the pull request — the workflow's own comment says it has published yesterday's date twice. The checklist said nothing about it.

ThePullRequestChecklist_AsksForWhatCiEnforces could not see the omission either: it anchors on ci.yml's step names, and this gate is not in ci.yml. It now reads auto-release.yml too and requires the release-only section to mention UTC.

Also: private workflow wording out of tracked files

Four tracked files described a particular two-machine development setup — "which happens on the other workstation", "verified on the laptop workstation", "not the build box". Rewritten to say what they mean ("needs the app actually running") in ArchitectureTests.cs, SidebarSelectionContractTests.cs, AudioPolicyConfigTests.cs and ARCHITECTURE.md. Zero occurrences remain in the tree; the same pass was already applied to 23 issue and PR bodies.

Verification

Mutation proof, both files restored byte-for-byte:

The second mutation is deliberately the case the old slice would have accepted. Baseline and post-restore 2 green, 0 red.

107 cases green across ArchitectureTests, SidebarSelectionContractTests, AudioPolicyConfigTests and ThemeServiceTests. Builds 0 errors / 0 warnings, dotnet format --verify-no-changes clean on both projects. test: — no version bump, no CHANGELOG entry, no release.