Coverage audit

This audit compares three lists:

1. The source inventory: routes, components and APIs read from the code.
2. Browser discovery: what was opened and exercised in a real browser.
3. Documentation: the feature and workflow pages in this package.

Summary#

DOCUMENTATION COVERAGE: 75%. Each feature scores 1 if it was fully exercised in the browser and documented with screenshots, 0.5 if it was only partly exercised, and 0 if it is documented from code alone: (40 + 13 × 0.5) / 62. Every feature has a documentation entry, but documentation without browser verification does not count as coverage.

Feature-by-feature#

Audit method#

1. Source inventory. A full read of app/: 28 user-facing page routes (plus 10 developer-only /dev previews) and 86 API routes, plus components/, lib/ and the migrations. It produced the feature list above, with source locations.
2. Browser crawl. Section by section, in a real Playwright-driven Chromium (gstack browse) at 1440 × 900, with a mobile pass at 390 × 844, using the synthetic QA account and the demonstration scenario.
3. Second pass for missed features. After the crawl, the source list was searched again for routes, tabs, dialogs, menus, exports, uploads, notifications and feature flags that the crawl had not touched. Items found and then walked in that pass:
   • the minutes' Dhivehi document language;
   • the /tasks and /m/<id>/minutes redirects;
   • print views (/print/<id>, ?version=approved);
   • the floating dictionary panel;
   • the Decision Register filters dialog;
   • the per-action More menu and delete confirmation;
   • the Library read view;
   • /join/<token>;
   • the signed-in 404.
4. Classification. Each screen is marked walked, shallow or skipped in the Screen atlas.

Remaining gaps, and why#

Notable UX findings (observed during the run)#

1. Stale minutes after an update: fixed 30 Sep 2026. After "Update · free" or a section regenerate, the minutes stayed on "Building the minutes…" until a reload, while refreshing every 6 s. The panel re-synced its status only when the server status changed, and an update goes ready → ready. It is now keyed on status and generation (lib/minutes/status-sync.ts, test/minutes-status-sync.test.ts). A related fix: tab switches wrote the address with replaceState(window.history.state, …), which Next.js does not follow, so refreshes fetched and restored a stale ?tab= (test/history-replace-state.test.ts). "Approve now" and series "Add actions" were re-tested and update on their own in ~2.5 s and ~1 s; the original notes came from checking too soon on a cold dev server.
2. Approved actions lose owner and deadline. The minutes extracted "Responsible: Mariyam Shifa · Deadline 9 October", but the approved action started "On nobody's list" with no due date.
3. Allowance counters disagree by one. The Assistant showed one fewer free question or draft than Settings → AI allowances.
4. Approval wording. On a 3-seat Team plan with no members yet, the approval drawer says "Your plan is for one person".
5. Summary keeps uncorrected names. After speakers were renamed, the Summary and Ask answers still said "Mariam"; only the minutes picked up the corrections.
6. Ask added an unsupported caveat, "the meeting transcript was truncated…", although the transcript was complete.
7. Grammar: "Ready to transcribe · 1 credits". Wording: "You gave it to you" in action history.
8. Deep link: the Actions tab is ?tab=todos; ?tab=actions silently opens the Summary.
9. Mobile: the floating dictionary button overlaps content, and the meeting tab bar cuts "Insights" until scrolled.
10. Accessibility: the Activity drawer did not close on Escape in the automated browser (its Close button works).

Verification checklist (performed before finishing)#

• [x] All 162 screenshot files exist and are listed in screenshots/manifest.tsv (checked by script: 0 unlisted, 0 missing).
• [x] Every Markdown image and link path resolves (checked by script, see below).
• [x] Portal links and gallery images resolve (same check, run on index.html).
• [x] No secrets in screenshots: the QA email, the one-time invitation link and the bank account number are redacted with overlays. No API keys, tokens or cookies appear on any page captured.
• [x] Sample data is clearly synthetic (see Sample data).
• [x] Features not exercised are marked NOT VERIFIED or "code only"; nothing unfinished is described as working (e.g. "talk-time per speaker — coming soon").