Phase log 85: Phase 56, TypeScript behaviour tests, coverage 71.9 to 99.5 percent, nineteen bugs fixed

Date: 2026-10-17 to 2026-10-18 (overnight). Scope: src/ts/** (what Codecov measures, flag typescript, vitest v8). The Rust side is phase log 83/84; this file does not touch it except the one resolver fix noted below.

Goal

Start: "Codecov says 61% of the code has tests, make it 100%." The maintainer then corrected the intent: the tests exist to check that the code matches the behaviour they expect, functionally and behaviourally, not to reach a number. So the work became: write tests that pin behaviour, let coverage rise as a side effect, and fix what the tests expose. Source was never changed just to raise a number, with one exception that is stated under Decisions (dead guards).

Current Progress

Final measurement, npm run test:coverage (exit 0, thresholds enforced):

• Test Files 171 passed (171), Tests 3264 passed (3264).
• npm run lint clean, npx tsc --noEmit clean, npm run format:check clean.
• vitest.config.ts floor raised from 90/82/89/92 to 99/95/99/99 (statements/branches/ functions/lines), just under the measured figures.
• Not run: npm run preflight (cargo fmt/clippy/test), ui-lab (Playwright), the Rust suite. The Rust fix below was verified with cargo test -p unv-app --lib only.
• Nothing is committed. Commit commands were handed over separately.

New test files (all under tests/)

Shell and grid: vault-wiring, vault-bundles, vault-desktop, vault-edge, app-shell, card-actions, card-display, card-interactions, grid-layouts, sidebar-trees. Config view: config-view-add, config-view-export, config-view-layouts, config-view-actions, chunk-cards, chunk-export-sparse, pick-file. Entry form: entry-form, entry-form-tools, modals-tail, save-entry-rules, cookie-session-form, form-generate, copy-menu (extended), dropdown-menu. Settings: settings-desktop, settings-order, settings-tabs, settings-garbage, settings-persistence, settings-backup, backup-restore-shapes. State and references: state-edge, persist-paths, reference-resolution, field-references, store-protocols, env-naming, env-link-modal (extended), project-slugs (extended). Tools and panes: tools-panes, tools-workflows, health-scan (extended), dev-tools, missing-markup, users-remote-contract, users-management, auth-panel-flows (extended), coverage-nodes-pane (extended), remote-connect (extended), unlock-flow (extended), permex-auth-more, small-tails, dialogs (extended), parsers (extended), bundle-export (extended), config-check (extended). Helpers: tests/app-boot.ts (boots the real app against a small vault, tracks and removes the document/window listeners it adds), tests/desktop-boot.ts (same, inside a faked Tauri shell).

The bugs that mattered

Ranked by what they did to a person. All fixed, each with a regression test whose comment says what the bug did.

1. A null in the stored settings stopped the app starting (state.ts, mergeStoredSettings). The first .slice on the replaced default threw in init(). Settings are plain JSON anyone can edit; values of the wrong kind are now dropped and the default kept.
2. Saving the entry form after a reload overwrote a neighbouring entry and kept its id (modals.ts). The form held only an index, so a peer edit or vault watch reload while it was open made the save land on whatever sat at that index. It now holds the entry by identity and refuses with "changed or deleted elsewhere" when it is gone (invariant 1).
3. Restoring a backup or JSON file without projects left the vault with no default project (import-export.ts).
4. Cyclic ${chunk:…} references returned their own ${…} text as the value, in TypeScript (chunk-ops.ts) and in Rust (unv-app/src/refs.rs). That literal reached exported configs.
5. Cancelling "add member to bundle" pulled an unlabelled entry into the bundle (vault.ts; label === undefined matched).
6. A public value could not be hidden again (render.ts): the public rule ran before the explicit reveal state, contradicting its own comment.
7. Group-by-type silently dropped cookie, composite and other kinds while the toolbar count still included them (render.ts).
8. The live name-collision warning in the entry form never fired (modals.ts): its draft held no values, so namesGeneratedBy returned nothing.
9. The nginx "create certificate" stub never joined the project it was made from (render.ts).
10. Stack export "Download" threw Cannot access 'dlText' before initialization (render.ts).
11. The eye button did nothing on its first click when the value started out visible (modals.ts, toggleReveal).
12. "All Secrets" stayed highlighted beside an environment filter (render.ts).
13. Toggle buttons rendered aria-pressed="" for false (render.ts); html prints false as nothing. Now String(...).
14. The recovery-package success message was overwritten by the summary at once (settings-panel.ts).
15. The lock-screen server picker showed the literal text   Local Vault (lock.ts).
16. The docker services card never showed its "No fields" hint (chunk-ops.ts, !subCards tested an array).
17. The add-2FA counter row stayed out of sync with the seed's kind (auth-panel.ts).
18. A closed dropdown or context menu still added its outside-click handler (modals.ts).
19. The Rust twin of bug 4 (counted once above as a file, listed here because it needs its own commit and a cargo test).

Decisions made

• Tests assert behaviour, not coverage. Hollow tests written early (assertion-less, or driven by /* v8 ignore */ hints) were deleted. Source changes made only for coverage were reverted.
• Dead guards were removed, not tested. Where an element comes from index.html (so the real DOM always has it) a guard such as if (!el) return or el?.value ?? '' is unreachable; those were converted to ! / plain access (about 60 sites in modals.ts, vault.ts, render.ts, lock.ts, settings-panel.ts and others). This is the one place source changed without a bug. It sits in the same commit as the bug fixes because the files overlap.
• Guards that tests rely on were kept: #custom-style (a <head> element, absent after a DOM swap), #search in resetViewState, #search-history, and everything in timeline.ts init. missing-markup.test.ts pins that the init functions tolerate missing markup.
• Removed dead code: the copy-chunk-raw handler and diffAndStashCopy in render.ts (no markup ever emitted the action), the env_file branch of chunkFieldCopyText, two unreachable yamlStr('') branches, and the unreachable tail of disambiguateNames.
• invoke in users.ts is exported so the remote users REST contract can be asserted command by command (users-remote-contract.test.ts).
• No ADR written: nothing here changes a design.

What Worked / What Didn't Work

Worked:

• Booting the real vault.ts (bootApp) and driving real buttons found most of the bugs; unit tests of helper functions found almost none.
• Table-driven tests for repetitive surfaces (21 add-chunk buttons, 7 exporters, 54 REST routes).
• Writing a "what would a person see" assertion first, then reading the code when it failed. Every one of bugs 1 to 16 came from a test whose first expectation was wrong in an interesting way.
• Reading coverage from coverage-final.json with a small script instead of the text table.

Did not work:

• /* v8 ignore */ hints: abandoned, they hide rather than test.
• Subset coverage runs: they miss tests in other files that cover a line, so gaps looked bigger than they were. Use the full run for decisions.
• Running the suite while another heavy job (cargo llvm-cov) held the 4 cores: timing tests flaked. Use --maxWorkers=2 --testTimeout=60000.

What is missing (coverage)

557 uncovered items remain (66 statements, 470 branches, 21 functions), in 30 files:

What they are, by kind (so the next person does not re-derive it):

• ?? / || fallbacks on dataset strings, API answers and optional fields where the markup or the server always supplies the value (most of the branch count).
• Tolerance for a missing DOM element in functions that would throw a few lines later anyway (the form functions in modals.ts).
• Document-teardown guards (typeof document === 'undefined' in ui-qol.ts): unreachable in jsdom by design.
• Unreachable else arms in settings-panel.ts drag-and-drop and a few from < 0 checks.
• Real but small behaviour left: lock.ts three anonymous functions (idle/visibility handlers), remote-panel.ts four (connect/test callbacks), state.ts eight (settings _apply callbacks, Settings.init migrations), users.ts two (permission save callbacks), settings-panel.ts three (drag handlers, archive picker).

Next Steps

For the maintainer (needs their hands or authority):

• Commit. Nothing was committed or pushed. Suggested groups: Rust resolver fix; src/ts fixes and dead-guard removal; tests; coverage floor. Foreign changes in the tree (Cargo.lock, other unv-app/unv-cli/unv-server/vault-core/src-tauri files, unv-cli/tests/*, src-tauri/tests/*, ADR-0159, SUMMARY.md) are not part of this work.
• Decide whether auth-panel.ts advance() should save through persist() (see Notes). It is a behaviour change, so it was left alone.
• Run npm run preflight and the Codecov upload once the branch exists; the figures above are local.

For whoever continues the tests:

1. lock.ts: drive the idle-timer and visibility-change handlers with fake timers (the three remaining functions).
2. remote-panel.ts: the connect-time TLS pin callbacks and the status refresh error branches.
3. state.ts Settings.init migrations: run init() with each unenverse-*-migrated flag unset.
4. users.ts: save callbacks of the permission editors (set_user_permissions, set_class_permissions) through the real editor, not the dispatcher.
5. settings-panel.ts: the drag arms that need a dataTransfer shaped like WebKitGTK's.
6. Only then consider the ?? fallbacks. Each one is a decision: test with absent data, or delete the fallback if the invariant guarantees the value.

Notes for whoever picks this up

• Measuring: npx vitest run --coverage --coverage.reporter=json --coverage.reporter=json-summary --coverage.reportsDirectory=/tmp/covfull --coverage.reportOnFailure=true --maxWorkers=3 --testTimeout=60000 --coverage.thresholds.lines=0 --coverage.thresholds.statements=0 --coverage.thresholds.branches=0 --coverage.thresholds.functions=0, then read /tmp/covfull/coverage-final.json (statement/branch maps) rather than the text table. The full run takes about 10 to 15 minutes; run it in the background.
• Flaky under load (pass alone): efficacy.test.ts, env-import-modal.test.ts (fixed flush to 120 ms), coverage-audit, icon-picker, modals.test.ts ("hashes on the hash tab"), entry-form-tools hash test (now waitFor). They are timing-sensitive, not wrong.
• Pitfalls found (candidates for .claude/rules/frontend-tests.md):
  • html renders false and null as an empty string, so a boolean ARIA attribute needs String(...).
  • A single-line <input> strips newlines from .value, so a multi-line paste into #f-key is silently one line.
  • resolveEditTarget needs the object the form was opened on; an index alone is not identity.
  • makeProject({ chunks: undefined }) really leaves the key undefined, which is how to test "project with no chunk list".
  • Dispatching click on a disabled button via .click() does nothing; use dispatchEvent(new MouseEvent('click')) to reach the handler's own guard.
• Findings not fixed:
  • auth-panel.ts advance() calls st.store.save(st.vault) directly instead of persist(), so it skips id normalisation, delta saves and conflict handling.
  • initAuthPanel throws when its markup is absent; the other panes tolerate it.
  • entryHasPayload counts a named variable with an empty value as a payload, so a typed entry with only suggested, empty variables saves as a hollow entry.
  • toggleReveal and masked() now agree, but the eye button's first click on a public field has no visible effect to explain to the user; consider a tooltip. coverage-push-2026-10.md` (goal, method, helper scripts, which guards must stay).

Status at exit (2026-10-18)

• Committed and pushed by the maintainer: branch raise-typescript-test-coverage, 19 commits (24af76b Rust resolver fix through 9989e9b coverage floor), cut from main at 061976a (version 0.55.1). PR #38 is open against main. Commit script: ~/Code/UnENVerse-commit-typescript-coverage.ps1.
• Parallel PR #37 (rust-behaviour-tests, phase log 83/84, release 0.56.0) is open too. The two branches do not touch the same files except possibly vitest.config.ts and the version number; unv-app/src/refs.rs differs only on this branch. Merge #37 first, then rebase #38 and rerun npm run test:coverage (the 0.56.0 bump and any Rust-side TypeScript-visible changes could move the figures).
• Not run after the final commit: CI. The figures in this file are from a local npm run test:coverage on the uncommitted tree that became these commits; the tree matched.
• Next action for the next session: check PR #38 CI, in particular the CI result job and the coverage thresholds (99/95/99/99), and rerun the load-sensitive tests listed in Notes if one fails on the runner.

1 line(s) about the maintainer's own machines were left out of this page.