jobtrackingapp/docs/cv-integration-qa/REVIEW_AND_MERGE.md

Review and merge — 2026-10-27

Review remains explicit: uncertain source is visible, missing records can be authored from owned source blocks, and no change is applied until the user approves it. Corrections are distinguished from literal source mappings. Current values and record labels are shown; existing targets refer to the compared profile revision. A stale response retains edits but clears target selection and confirmation, requiring renewed comparison.

Object mappings with no existing target deliberately add a record; a selected target updates only approved supplied fields. Scalar groups append using existing canonical normalization. Contact writes affect selected leaves only. Rejected originals are absent from applied changes; mapping-only application is supported. Similar titles alone do not establish a duplicate. Incomplete identities stay separate unless a user selects the destination. Missing fields never clear existing fields.

New mappings retain submitted order. Previously sorting randomly generated client IDs reordered jobs. Replay hashes now preserve mapping order, and exact historical sorted-hash receipts remain replayable only when their stored mappings have exactly the same order. A reordered request conflicts rather than returning a false successful acknowledgement.

Verification

• CvImportMergeScenarioTests: exact/revised/older/overlapping imports, normalized dates, similar distinct employers, incomplete identity, target changes, selective correction/rejection, duplicate field/client-ID checks, source retention, detail deduplication and submitted ordering.
• Controller regression: legacy sorted receipt accepts the identical ordered request and rejects reordered mappings without another history write.
• RED ordering and replay failures retained in .artifacts/cv-integration-qa/mapping-order-*.log; successor merge-order-green-02.log: 126 focused tests passed, including parser/mapping/controller suites. This overlaps other focused reports.
• Actual Docling browser run real-browser-01.json: 6 passed, zero retries. Two existing journeys plus four new integration journeys; do not count the two existing cases twice. New cases cover EN/light desktop, Bokmål/dark mobile, tablet education authoring and repeated overlapping roles.
• Real UI tests inspect source, correct/map jobs or education, approve and verify canonical readback, compare existing targets, exercise a two-session stale write and reconfirmation, and preserve an unrelated variant/canonical data while customizing a CV and exporting PDF. Repeated-role reimport/replay does not add records.

This proves the exercised field groups and interactions, not every catalog field in every viewport. The wider catalog has backend/provider coverage. Automated accessibility assertions are separate from human assistive-technology acceptance. Original PDF tables with unresolved relationships still require user assignment; no automatic uncertain merge is claimed.