cesnimda
5.7 KiB
CV Integration QA: accessibility and manual qualification
27 October 2026. This is a bounded local evidence handoff for CV-QA-08, not human acceptance or a WCAG conformance statement. Authority: PLAN.md, architecture and coverage, and the unchanged redesign acceptance requirements.
Verified automation
.artifacts/cv-integration-qa/real-browser-01.json records six passing Chromium cases, zero skipped, unexpected or flaky cases. Two are the existing real Docling journeys; four are the new job-tracker-ui/import-tests/cv-integration-qa.spec.ts journeys. These four are included in six, not additional passes. The report began at 2026-10-26T23:18:21.414Z (27 October locally). The completed full regression separately passes364 ordinary and53 configured diagnostic/contract cases, including browser zoom, reflow, focus and reduced-motion checks. An additional long real-import case passes but adds no axe scan. None of these counts establishes human acceptance.
| New real Docling scenario | Locale / theme / CSS viewport width | Exercised behavior |
|---|---|---|
| Employer-first ambiguous DOCX | English / light / 1440 | Explicit source mapping, current-target comparison, 409 preservation and deliberate reconfirmation, editor/PDF, frozen application and letter preservation |
| Title-first ambiguous PDF | Bokmål / dark / 390 | Explicit source mapping, mobile review panel, durable profile, scoped editor/PDF |
| Ambiguous education DOCX | English / light / 768 | Explicit institution/qualification mapping, durable profile, scoped editor/PDF |
| Repeated overlapping roles DOCX | English / dark / 768 | Separate source occurrences, extracted review, same-source reimport, stable record IDs, scoped editor/PDF |
Each new scenario sets prefers-reduced-motion: reduce. The helper focuses selected buttons and activates them with Enter for add-mapping, confirmation and apply actions. This proves those sampled keyboard activations, not a complete Tab/Shift+Tab traversal, keyboard-only journey or focus-order assessment. Forms are also driven through Playwright locator actions.
The four new axe scans include only #cv-import-{runId}-review-panel, with WCAG 2 A/AA and 2.1 A/AA tags. The two retained import cases analyze the loaded page without that include restriction. All six attached scans record zero violations and one incomplete color-contrast result each, classified serious by axe. These incomplete results require inspection; zero violations does not establish contrast compliance. The new scans do not cover the whole application, source PDF iframe or physical documents.
Screenshots of mapped/extracted review and tailored preview are retained under .artifacts/cv-integration-qa/real-browser-01-results/ and attached in the report. The agent inspected representative desktop/mobile review screenshots and first/last rendered pages of all three long imported PDF templates. This is agent visual inspection, not human acceptance. Independent PDF oracles and their typography limitations are documented in PDF_FIDELITY.md; printing remains untested.
Open human and device checks
| Status | Required check | Owner action / acceptance evidence |
|---|---|---|
| OPEN | Contrast, including every axe incomplete result, focus indicators and disabled/error/success states | Reviewer inspects rendered foreground/background pairs in both themes and records affected control, measured ratio and disposition. |
| OPEN | Complete keyboard traversal and recovery | Human traverses source/review tabs, selects, edits, removes, confirms, applies and recovers from conflict without a pointer; records focus order, visible focus and focus after updates/errors. |
| OPEN | Assistive technology | Test representative screen reader/browser combinations for names, legends, status/error announcements, source evidence, mobile tabs and confirmation state. |
| OPEN | Physical mobile/tablet use and text reflow | Check touch targets, native selects, virtual keyboard, landscape, zoom/text enlargement and long content on actual devices. CSS viewport widths are not physical-device evidence. |
| OPEN | English and native Bokmål comprehension | A qualified reader checks wording, ambiguity/correction warnings, source/current distinctions and stale-target reconfirmation. Locale automation is not linguistic approval. |
| OPEN | Reduced-motion comfort and visual review | Inspect the whole journey, motion, loading and transitions under the actual OS preference; screenshots disable animations and do not prove motion suitability. |
| OPEN | Physical rendering and print | Independently inspect downloaded CVs, long/Unicode content, page boundaries, fonts, clipping and print results across supported templates/platforms. Preserve existing PDF oracles as separate evidence. |
| OPEN | Human end-to-end acceptance | Owner completes representative synthetic import-to-profile-to-CV/application tasks and records whether assignment and preservation are understandable and correct. |
Limits and next action
Real Docling parsing and local browser automation do not qualify live CV-model inference; no live inference was run. The independent static code review found no confirmed new blocker in the scoped changes, but retained unlabelled pipe order remains positional: reversed roles need a recognized table header or explicit mapping. This is a preserved contract limitation, not broad semantic accuracy evidence.
Owner: retain the reconciled report/screenshots and complete and record the open checks. Keep all 57 preservation groups and 33 routes in scope. Automated scoped checks support continued development assessment; human, device, assistive-technology and print acceptance remain OPEN, and production approval is not established.