Skip to content

Implement D79: deterministic structure skeleton + bottom-up summaries consumed by E1/E2 #165

Description

@fazpu

Owner-approved 2026-07-27 ("this suggestion is really good and should be done - we need the system to scale"). Design: PR #164 — D79 in decisions.md + plan/designs/e0_files_design.md §4.1.

Scope:

  1. Deterministic markdown-heading parser for the section skeleton; LLM fallback for structureless documents returns searchable heading strings, never char offsets. §4 output contract unchanged (document_sections on the block grid, pageindex.json sidecar).
  2. Bottom-up summaries: flash-class summary seat (D70 port binding), one bounded call per leaf section, parents composed from child summaries; parallel; content-hash cached per section.
  3. Consumption: summaries into the E1 context-prefix input and the E2 bundle as the D31 "section path + summary" element (target + ancestor summaries); added_context gains a summary tag with grounding-gate verification; extractor version bump. Closes the summary half of E2 context bundle diverges from the D31 design: section summary and entity hints are missing #163.
  4. Provenance: split structurer_version into skeleton version + summary seat/version.

Sequencing: after #161 (loss ledger) merges, so the bundle change lands with measurement in place. Judge summary quality only where consumed (#150 canaries).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions