v0.4.2: Reviewed questionnaires and source evidence

• Restore Northern Ireland's literal argument-coder labels. CSV type inference
  had collapsed 176 comma-separated code sets; source whitespace is also retained.
  Respondent identities, missing slots and all knowledge outputs are unchanged.

• Preserve Denmark's four original questionnaires and readable PDF companions;
  record the later follow-up's verified telephone mode without changing dates,
  scores or cohorts. Preserve its wave crosswalk, two California index drafts
  and Northern Ireland's version 16 coding guide in their poll folders.

• Allow the full CI build and numerical audit up to 60 minutes. The previous
  30-minute limit canceled a run after tests and lint passed, before its final
  metadata verification could finish.

• Use reviewed whole-questionnaire evidence in knowledge scoring. Preserve
  observed blank quizzes as zero; unavailable questionnaires have missing scores.

• Retain verified attendance before inferring nonattendance from absent exits,
  including Australia's nine attendees without exit questionnaires and two
  UK Health attendees without completed post-event questionnaires.

• Preserve Australia's original combined DP codebook and its readable PDF,
  distinguishing it from the separate constitutional-referendum codebook.

• Export 829 source-level attitude definitions and 1,039,910 response rows for
  Denmark, Vermont, Marousi and America in One Room 2024. Retain all source
  people, original identifiers, labels, units and documented interview phases.

• Exclude 372 reviewed attitude nonanswers and 40 invalid Vermont responses
  from numeric values. Preserve 271 Marousi responses with unverified scales
  as unclassified raw responses, with missing numeric values.

• Preserve Marousi's original identifiers separately for telephone, arrival
  and exit while retaining the authored row bridge and its 16 unresolved
  departure-ID conflicts.

• Apply reviewed full-form masks across ten polls and calculate group knowledge
  means and learning opportunity from observed peers. Preserve BTP General
  Election's 299-person group cohort before its 248-person analytical selection.

• Declare complete questionnaire-presence dependencies and use registered source
  column order. Individual recodes support reordered inputs and local subsets
  without changing reviewed values or weakening production identity checks.

• Retain 2,299,630 additional raw questionnaire responses used by the presence
  rules; every existing raw response and reviewed measure value is unchanged.

• Apply 152 codebook-backed CPL nonanswer declarations through a shared
  field-specific registry. Preserve substantive Other codes and dollar amounts;
  raw values and reviewed scores remain unchanged.

Validation: full local tests and lint; independent source, numerical and document reviews; all 35 respondent, analysis and aggregate output files rebuilt byte-for-byte without Git history or vault. Exact-commit CI is required before publication.

Australia combined none/DK and San Mateo empty exit retain existing scoring pending consultation. Missing sources and measurement decisions are documented in docs/poll-issues.md.