Pair attached headings with unique exact legacy body spans while preserving source identities and recording absorbed skipped headings. Keep mid-paragraph ordinals in prose, including indented-wrap and real code-boundary distinctions. Add continuation provenance, separate preprocessor/harness fingerprints, legacy detector-delta regressions, and historical measurement labels.
All 18 suites pass. Independent Sol review and verified 38-source comparisons preserve headline scoring metrics and expose the corrected diagnostic change buckets. Follows Claude's final #290 findings.
Fixes#293.
Co-authored-by: Codex <codex@openai.com>
Apply the ordinary 50–400 source-word eligibility rule to standalone and detached headings, retaining short-heading diagnostics and oversized-unit reasons. Cover eligible ATX/setext headings and consecutive headings so source coverage does not depend on attachment.
Co-authored-by: Codex <codex@openai.com>
Separate structural normalization from paragraph selection. Preserve document tokens, keep eligible bodies at the 400-word boundary, and retain detector-visible headings, quotes, lists, and code boundaries. Add verified-source accounting, opt-in JSONL provenance, and a fixed-detector comparison with legacy preparation.
Carries forward Francisco Quinteros's heading/body boundary fix and regression intent from #180. Implements the reviewed repair scope in #288 and design refinements to #289.
Co-authored-by: Francisco Quinteros <155692279+JavierQuinan@users.noreply.github.com>
Co-authored-by: Codex <codex@openai.com>
Conor Bronsdon