jobtrackingapp/docs/pre-phase10-ux/CV_QA_HANDOFF.md

CV-QA / PRE-P10-UX handoff — 2026-10-29

Read cumulative acceptance, evidence summary, local runtime acceptance, and the historical A/B/C/D checkpoint reports before resuming. Development acceptance is conditional; production remains NO-GO. This handoff does not authorise Phase 10 implementation.

Next bounded QA work

1. Review the revised S01 PDF failures against the unchanged 2 mm coordinate oracle on both platforms. Preserve failed evidence and the historical v1 renderer. Obtain an explicit measured disposition; do not replace the oracle with the separate ink check.
2. Perform keyboard/screen-reader, contrast, physical desktop/tablet/mobile, EN/Bokmål language and actual printing reviews. Exercise blank creation, undo/redo, rendered gallery, matching letters, photo crop, dashboard keyboard drag/resize and admin polling. Browser automation does not complete these human gates.
3. Qualify a CV interpretation model only in an approved isolated environment with immutable verified artifacts. Bind provider/runtime/model hashes, prompts, schema and source fixtures. Run the same frozen PDF/DOCX corpus, distinguish development examples, and measure typed precision/recall, missing facts, record association, unsupported claims, provenance and user review burden. Compare deterministic and model-assisted results independently. Test injection, malformed output, unavailable service, timeout, cancellation, repeated requests and unsupported/scanned layouts. Do not infer CV factual accuracy from synthetic strategy/rewrite smoke tests.
4. Qualify the inherited MySqlConnector advisory remediation and rerun MariaDB/RD21 rollback, stale writes, retries and ownership. Recheck all generated schemas/resources are removed. Preserve unrelated profile details and frozen CV revisions.
5. Before any production decision, separately obtain operator evidence for filesystem integrity, off-disk backup/restore, native parser isolation/resource policy, cache/artifact security, deployed provider configuration and exact deployed identity. Never use the development compose scripts against production.

Evidence and reproduction

The A–D application source is 3255da9; broad regression freeze 994bf8f adds test-only compatibility corrections. Subsequent caa0c6a corrects the strategy JSON boundary, with a new full backend/Python run and actual local worker proof. Its Linux API image is sha256:14adb1e05757575e7c9d50c8f504f8209ebbf7f9b5f870daef0113d9dfe7dd49. The local frontend is a separate /api build, not the localhost:5302 browser-test export. The latter was verified and archived before rebuilding. Raw evidence paths and hashes are in the machine-readable summary and local runtime report. Do not attribute earlier broad browser/PDF evidence to a later all-suite run.

Use the repository-supported runners, explicit managed worktree, repository .NET SDK and RTK wrapper. Serialize native build/browser/MariaDB lanes. Preserve nonzero exits and compare exact discovered test identities; do not sum focused reruns into overlapping full totals. The original Windows/Linux PDF oracle remains authoritative for its own result.

All 57 preservation criteria remain carried forward, including the original PARTIAL/DEFERRED qualifications. A route being present is not proof of complete functionality. The next workstream may assess cleanup candidates and their risk; this task stops after publishing this handoff and does not start that implementation.

Required publication action

The Git server stopped responding on port 443 during the final correction/evidence publication. The last remotely verified checkpoint is 994bf8f; caa0c6a and the final evidence commit are local until connectivity returns. Once available, run a normal push of test/pre-phase10-ux-acceptance and compare git rev-parse HEAD with git ls-remote origin refs/heads/test/pre-phase10-ux-acceptance. Do not force push, merge or deploy. This is the remaining owner/operator action; it does not require rerunning already source-bound evidence unless application source changes.