chore(tests): delete 14 source-text tests that no defect can fail

[PathGao]

What this measures

node --test cannot import a .svelte file, so anything asserted about a component can only be a regex over its source text. 19 files were nominated for deletion on the theory that such a test detects a rename and nothing else. I did not take the list on faith: every file was measured by injecting a real defect into its own stated subject and running the whole suite.

14 are deleted. 5 are kept — each was measured to catch a real defect that nothing else catches. Four assertions from inside the deleted files were salvaged.

Deleted: the file stayed green with a real defect live

Each row is a defect injected into production code, with the whole suite run against it. "Also caught by" is the file that went red instead — where that column is empty, nothing in the suite noticed at all.

Deleted file	Defect injected	File red?	Also caught by
previewScrollSync	scroll-sync threshold 5 → 100000 (sync dead)	green	—
previewScrollSync	sync ratio hard-coded to 0	green	—
findCollapsedMatches	revealFoldsAround made a no-op	green	—
findCollapsedMatches	foldToggleFor always returns null	green	—
toolbarCustomizationWiring	reorderEditorToolbarTool made a no-op	green	—
toolbarCustomizationWiring	reorderTitlebarToolbarAction made a no-op	green	—
interruptedSessionRestore	restore breadcrumb never persisted	green	sessionRestoreResilience
interruptedSessionRestore	explicit exit stops clearing the snapshot	red	windowStateRestore
taskToggleMemory	checkbox toggle writes the wrong line (=== → >=)	green	truncatedBufferGuard
taskToggleMemory	rendered task boxes stay disabled	red	renderProtocol
reloadOpenToolbar	reload calls canCloseTab and ignores the answer	green	—
reloadOpenToolbar	the fmt-bold toolbar action does nothing	green	—
windowsPdfExport	await dropped from invoke('export_pdf_windows')	green	—
batchCloseConfirmation	Close Others also closes the tab it was invoked on	green	—
frontMatterDisclosure	the front-matter tag remove button is inert	green	—
viewerDisposal	one of seven isDisposed guards removed	green	—
previewRenderRevision	the stale-render guard's || becomes &&	green	—
tocEditorJump	the ToC jump selects line 1 instead of the heading	green	—
openMultipleFiles	only the first selected document is opened	red	—
tabContextMenuIsolation	overlay left-click no longer dismisses the menu	red	—
editorOptionWiring*	minimap and wordWrap settings ignored	green	—

The rows marked red need the qualifier that decides this PR. In each of them my edit landed on the exact line the file greps, so the file failed because a string moved, not because behaviour changed. That is the same mechanism that turns 2 tests red for renaming revealFoldsAround while 553 stay green for making it a no-op. It is detection of an edit, not of a defect — and svelte-check already refuses an inconsistent rename, which is the only edit of that shape that is ever a bug.

Where the injected defect avoided the greped line (viewerDisposal, previewRenderRevision, batchCloseConfirmation, frontMatterDisclosure, tocEditorJump, reloadOpenToolbar, findCollapsedMatches), the file was green. That is every file in this list.

* editorOptionWiring appears here for context but is kept — see below.

Kept: 5 of the 19, each with a defect only it catches

These were on the list. They should not have been. In each case the file caught something that is not a spelling pin, and nothing else in the suite caught it.

Kept file	Defect it caught alone	Why it is not a spelling pin
editorOptionWiring	one of the two renderWhitespace sites regresses to "trailing"	the assertion is count(...) === 2, so it catches creation and updateOptions disagreeing — a class of bug, not a string
printFindHighlight	the print rule's background-color: transparent becomes display: none	resolves the CSS cascade between styles.css and a :global() rule in FindBar.svelte; no compiler does this
foldLayout	styles.css fold wrapper reverts to grid-template-rows: 1fr	couples foldLayout.ts (writes --fold-content-height) to styles.css (consumes it) across files
scrollSyncInput	a cursor-position listener added to the scroll-sync effect	a negative assertion catching newly added code, which no rename can trip
windowTags	windowTag dropped from the v2 window snapshot; pinned forced true	the only check that the tag survives a serialize/restore round trip

The table's premise for windowTags — "Rust pinned-tag write made a no-op" — reproduces: making save_pinned_tag_at return Ok(()) leaves all 562 JS tests green. But cargo test catches it in three tests (window_runtime::tests), and the JS file's remaining value is the frontend half, which nothing else covers. Deleting it would open a real hole in a data-persistence path.

Salvaged

Four assertions pinned something no compiler can see. They moved rather than died, and each was re-verified by mutation after the move.

From	Assertion	To	Verified by
windowsPdfExport	every invoke('…pdf…') name in export.ts is registered in generate_handler! and defined as a #[tauri::command]	macosPdfExport	renaming export_pdf_windows on the Rust side only → red
toolbarCustomizationWiring	11 interactive labels are translated in every locale; 11 window-organization labels are translated in zh-CN	i18nCoverage	deleting the zh-CN menu.mergeAllWindows entry → red
findCollapsedMatches	FOLD_TRANSITION_MS outlasts the styles.css height transition	foldLayout	lengthening the CSS transition to 0.6s → red
tabContextMenuIsolation	create_transfer_window is async fn	windowOrganization	dropping async → red

The two i18n tests were the only behavioural code in toolbarCustomizationWiring — they import the dictionary and look keys up in it. i18nCoverage deliberately does not enforce per-locale completeness (19 locales are >100 keys short and failing on that would block every English string), so these keys needed to survive as the named exception: they label icon-only controls where the word carries the whole meaning.

The PDF salvage reads the command names out of export.ts instead of restating them, so a renamed command is checked at its new name rather than quietly passing.

Coverage holes this PR does not close

Deleting these files removes no detection, but it does remove the appearance of it. Four subjects now have no test at all, and each needs an executable one:

• Preview/editor scroll sync. getScrollSyncPositionFromPixels and getScrollTopForSyncPosition are pure functions of three numbers, currently declared inside MarkdownViewer.svelte and Editor.svelte where nothing can import them. Moving the pair to src/lib/utils/scrollSync.ts would make the whole subject unit-testable, including both defects above.
• revealFoldsAround / foldToggleFor in FindBar.svelte. These are DOM walks; foldMeasurementBatching.test.ts already drives observeFoldLayout through a recording stand-in for the DOM, and the same technique applies here.
• exportAsPdf. export.ts is importable and already covered behaviourally for HTML export (exportFoldParity, exportRichContent). The dropped await above would be caught by an equivalent test with a stubbed invoke/save.
• reloadFromDisk's dirty-buffer guard. documentSession.canCloseTab is covered behaviourally by truncatedBufferGuard; the reload path that calls it is not, because it lives in the component.

Verification

• npm test 524 pass / 0 fail (was 562); npm run check 0 errors; npm run build clean; cargo test 144 pass.
• 199 assertions removed with the deleted files; the salvage adds 5 tests back.
• Every mutation the suite caught before, it still catches. Re-run after deletion: breadcrumb → sessionRestoreResilience; snapshot discard → windowStateRestore; task marker → renderProtocol; fold heights → foldMeasurementBatching; fold styles → foldLayout; window tag → windowTags; whitespace → editorOptionWiring; cursor listener → scrollSyncInput; print highlight → printFindHighlight.
• The three control mutations all still caught: sanitizer stops forbidding <style> → exportSanitize + previewSanitize; lossy-decode save guard inverted → checkedReadMigration + externalChangeReload + pathIdentityCaseFolding + truncatedBufferGuard; checkbox toggle writes the wrong line → truncatedBufferGuard.
• scripts/sourceTree.ts is untouched — none of the deleted files imported it (its only consumers are previewSanitize, renderPipelineConvention and singleImplementationConvention).
• package.json: test:settings-scroll is removed (both its files are gone); test:workflows and test:frontmatter drop the deleted entries.

🤖 Generated with Claude Code