test(scripts): re-anchor 12 assertions that could not fail

[PathGao]

Twelve assertions in twelve test files survive the defect they name. Each was reproduced first: the mutation applied to a clean worktree, npm test run, 584/584 green, source restored. None of them is the failure #440 fixed — those were anchors that were missing, so the slice degenerated to "". These are anchors that are too far apart, values compared against the constant they were read from, and assertions with no reachable subject.

These tests are recent, and they were added by the same series of changes that is now fixing them. Ten of the twelve files did not exist at v2.6.14; only editorToolbar.test.ts and windowClosePerTab.test.ts predate it. This is not inherited debt.

Based on df80968 (#447). npm test 586/586, npm run check 638 files / 0 errors, npm run build clean. No Rust touched.

The one that was live

mermaidPrintTheme.test.ts exists so that PDF export renders diagrams light — Mermaid bakes theme colours into the SVG as attributes, so a dark preview exports dark boxes with dark text onto white paper. The file asserted assert.deepEqual(mermaid.themes, [MERMAID_PRINT_THEME]), and MERMAID_PRINT_THEME is the constant renderDiagramsForPrint reads.

-export const MERMAID_PRINT_THEME = 'neutral';
+export const MERMAID_PRINT_THEME = 'dark';

584 tests, 584 passing. Every PDF export re-themes every diagram dark, and the file whose whole purpose is the print theme has nothing to say about it.

Making the implementation stop using the constant is caught, so the call sequencing was real — only the value was unpinned. The assertion now spells 'neutral', and MERMAID_PRINT_THEME was exported for nothing else, so the export goes with it.

Shape 1 — value pins that pin nothing

site	what it said	what it holds now
mermaidPrintTheme.test.ts:103,108,114	deepEqual(mermaid.themes, [MERMAID_PRINT_THEME])	['neutral'] — the literal is the claim
editorToolbar.test.ts:14-26,41	every expected value derived from DEFAULT_EDITOR_TOOLBAR_ORDER, which is EDITOR_TOOLBAR_TOOLS.map(t => t.id)	the id list as literals, plus each tool's label / name / shortcut

Deleting the Underline tool from the editor toolbar catalogue outright left the whole file green: expectations written in terms of a derived constant shrink with it. titlebarToolbar.test.ts has the same derived shape but also asserts literal id lists, so the same deletion there reddens three tests; editorToolbar.test.ts now does too. The added test also covers name, label, group and shortcut, which nothing tested — a tool that silently loses its accelerator still renders.

Shape 2 — discriminators neutralised

exportRichContent.test.ts:299 — "a document with no math carries no KaTeX fonts at all" called inlineKatexFontFaces(css, new Set(), new Map()). The empty byte map deletes every face through the "bytes could not be read" branch, so the empty family set decides nothing. Deleting the family filter from exportFonts.ts:171 was green across the whole suite, integration assertions included. The test now supplies every face's bytes, so the family filter is the only thing left that can drop them.

previewAnchorRestore.test.ts:327 — /import \{[\s\S]*findAnchorElement[\s\S]*\} from '\.\/utils\/previewAnchor\.js'/. Both gaps are unbounded, so the match starts at the first import in the file and runs past every brace between there and line 46. Forking findAnchorElement / getAnchorScrollTop into another module, leaving the rest of the import in place, was green. Now read out of the one import statement ([^}]*) and checked as a specifier list.

editorPdfExport.test.ts:379-380 — the test is named "a failed refresh does not pass off a stale export as a fresh one" and asserted /catch \(error\)/ and /addToast\(/.

-addToast('Exported PDF may not include the latest edits', 'warning');
+addToast('Export refreshed', 'success');

Green, both regexes still matching, with the claim in the test's name inverted. The export deliberately still goes ahead on a failed refresh — printing the stale DOM beats not printing — so the entire contract lives in the severity. That is what is asserted now: exactly one toast in the catch, severity warning.

Shape 3 — over-fitted spellings

Two doesNotMatch guards pinned the exact one-line spelling a fixed defect happened to be written in. Both defects were reintroduced in ordinary brace / for…of spelling; both green.

liveModeWatchedPath.test.ts:26 — /isLiveMode\(\)\) invoke\('watch_file'/ against documentSession.svelte.ts, where neither token occurs any more. #296 removed the isLiveMode option from that module entirely, so the guard has been reading a file that cannot contain its pattern. Re-anchored to who may own the watcher: filesMatching(readSourceFiles('src'), /invoke\(\s*'watch_file'/) must be ['src/lib/MarkdownViewer.svelte'], and callbackBodies(viewer, '$effect') must contain exactly one effect that watches — which is also where the matching unwatch_file has to live.

windowClosePerTab.test.ts:25 — /tabManager\.tabs\.forEach\(\(t\) => \(t\.isDirty = false\)\)/. Now /\.isDirty\s*=\s*false/ over the close handler: the assignment, not the loop it is written in. (This is one of the two files that predate v2.6.14.)

Deleted rather than repaired

Five sites have no reachable subject. Each is dominated by an assertion that already covers the ground, named in a comment where the dead one was.

site	why it cannot fail	what covers it
exportSanitize.test.ts:56	assert.match(POC, /onerror=/) — POC is a template literal three lines above	the config assertions below it; uriIsAllowed('x') runs the real regexp and stays
i18nCoverage.test.ts:453	assert.ok(total >= 0) on a sum of array lengths	:249 every key the source asks for exists in English, :267 no language defines a key English does not have
settingsPersistence.test.ts:195	new SettingsStore() instanceof SettingsStore — a constructor that throws fails the line above, one that returns fails nothing	the effect-count assertions two lines down, which are the point of the test
renderProtocol.test.ts:252-256	underscore counts of two strings :242 already asserted strictly equal	:242-246
scrollSync.test.ts:52	(200/1000)*600 === 120 over the file's own two fixture constants	:54-55, on numbers the mapping returned

The renderProtocol case was checked rather than argued: making processDisplayMathBlocks strip underscores from data-math-source fails at line 242, so line 252 never runs.

scrollSync.test.ts:52's 120 is genuinely useful documentation — it is the number the whole two-section design exists to avoid — so it stays, in the comment that already stated it.

Reported as unreachable, and is not

exportSanitize.test.ts:68 — assert.equal(MARKDOWN_SANITIZE_CONFIG.ALLOWED_URI_REGEXP, ALLOWED_MARKDOWN_URI_REGEXP) compares the field against the constant it is assigned from one line earlier in sanitize.ts, and the two are the same object reference. But the assignment is reachable:

-	ALLOWED_URI_REGEXP: ALLOWED_MARKDOWN_URI_REGEXP,
+	ALLOWED_URI_REGEXP: /^[\s\S]*$/,

fails it, and previewSanitize.test.ts with it. Left as it stands.

Mutation check

Each mutation applied to df80968, the suite run, the source restored; then re-applied against this branch. "before" is the full suite (584 tests at df80968), "after" is the file under test.

#	site	mutation	before	after
1	mermaidPrintTheme:103,108,114	MERMAID_PRINT_THEME 'neutral' → 'dark'	584/584 green	RED — exporting re-renders with the print theme…: actual: [ 'dark' ], expected: [ 'neutral' ]
2	mermaidPrintTheme:187-189	divert only MarkdownViewer.svelte:1072 (preview) to mermaidTheme: 'dark', leaving the HTML export at 1886	584/584 green	RED — the PDF export wraps the print render…: actual: [ "'dark'" ], expected: [], every diagram render must resolve the theme the way the print restore does (currentMermaidTheme())
3a	mermaidPrintTheme:152	delete if (diagrams.length === 0) return () => {};	—	RED — the helper is inert…: actual: [ 'neutral' ], expected: [], nothing to re-render must not re-theme mermaid
3b	mermaidPrintTheme:152	delete if (!root || !mermaid) return () => {};	—	RED — the helper is inert…: TypeError: Cannot read properties of null (reading 'initialize')
4	editorToolbar:14-26,41	delete the Underline tool from EDITOR_TOOLBAR_TOOLS	584/584 green	RED (3) — the default order is the whole tool catalogue…, each tool carries the label, name and shortcut…, getVisibleEditorToolbarTools applies saved order and hidden ids
5	exportRichContent:299	delete if (!usedFamilies.has(face.family)) continue; from exportFonts.ts:171	584/584 green	RED — a document with no math carries no KaTeX fonts at all: input matched /@font-face/
6	previewAnchorRestore:327	fork findAnchorElement / getAnchorScrollTop into a second module, leave the other two specifiers	584/584 green	RED — the viewer restores through the measured resolver: actual: [ 'findAnchorElement', 'getAnchorScrollTop' ], expected: []
7	editorPdfExport:379-380	warning toast → addToast('Export refreshed', 'success')	584/584 green	RED — a failed refresh does not pass off a stale export…: actual: [ 'success' ], expected: [ 'warning' ]
8	liveModeWatchedPath:26	documentSession.loadMarkdown re-issues invoke('watch_file', …) in an ordinary if (…) { … }	584/584 green	RED — Live Mode follows the active file…: actual: [ 'src/lib/MarkdownViewer.svelte', 'src/lib/sessions/documentSession.svelte.ts' ], expected: [ 'src/lib/MarkdownViewer.svelte' ]
9	windowClosePerTab:25	for (const t of tabManager.tabs) t.isDirty = false; at the top of the dirty branch	584/584 green	RED — the aggregate unsaved-files modal is gone from the close handler
10	renderProtocol:252-256	data-math-source written with underscores stripped	—	RED at :242, never reaching :252 — which is why :252-256 is deleted
11	exportSanitize:68	ALLOWED_URI_REGEXP: /^[\s\S]*$/	RED (2) — does not reproduce	unchanged

One honest gap. The survey's mutation for site 3 was make the early-return closure write a global; that is still green here, and the rewritten test does not claim otherwise. Writing an unrelated global is not a defect shape this helper has. What the test now claims — and what 3a and 3b prove — is that the handle the export's finally calls re-themes nothing and touches no DOM.

Not covered

truncatedBufferGuard, viewModeWithoutSaving, previewSanitize, windowOrganization, recentFilesMultiWindow and editorContextMenuI18n are untouched here — six sites where an unprotected guard, rather than a dead assertion, is the finding.

🤖 Generated with Claude Code