[coderback]

Summary

The first part of Phase 2 (CVs and cover letters): every tailored CV and cover letter is checked against the candidate's own data and the job ad, both while it's generated and in the job workspace.

• Document check (api/app/doc_check.py, no model calls). For a CV or letter it reports:
  • the ad's technical terms, split into those the document shows, those the candidate has but it leaves out, and real gaps (never added);
  • numbers, technologies, employers and project names the persona doesn't contain, field by field (a letter may also quote the ad);
  • words recruiters read as machine-written;
  • the rendered PDF read back as plain text, as an ATS reads it: name, email and phone, section headings in order, every skill, role dates, page count.
• POST /documents/check returns that report for the signed-in user.
• Generation:
  • A tailored CV gets the ad's terms that are among the candidate's own skills added to the right skills row (profile spelling only).
  • When the check flags a claim or a word, one targeted revision runs on the CV text or the letter's paragraphs. It is kept only if it clears findings without adding new ones.
• Job workspace: a Checks strip on the CV and cover-letter steps re-runs a moment after each edit. Its details list the keywords, each flagged claim and where it is, and every ATS read-back check.

Real-model run

Five stored jobs, against the real persona and model:

• CVs: all 5 came back with no unsupported numbers, tools, employers or projects, on one page, and reading back cleanly from the PDF.
• Cover letters: 2 of 5 had one flagged claim; the revision cleared both.

Testing

• API suite: 680 passed. The new tests cover:
  • term and number matching, including ordinary words like "rest", years, dates and ranges;
  • the keyword split and the fact and style findings;
  • the endpoint;
  • the skills top-up;
  • revision kept, dropped when it invents something new, and dropped when the reply has the wrong shape;
  • letter revision.
• The PDF read-back tests run in the API image, where WeasyPrint loads: 24 passed.
• Frontend build and lint pass. The e2e mocks can now stream SSE, and a new smoke test walks the workspace from fit analysis to a checked CV: 10 passed.