E2 context bundle diverges from the D31 design: section summary and entity hints are missing

[fazpu]

The D31 bundle table (plan/designs/e2_e3_claims_relations_design.md §3.1) says the extractor receives, per target chunk:

1. Document header (title, date, source, language) — implemented
2. PageIndex section path + summary — PARTIAL: _bundle_text (workers/e2.py) renders SECTION: path …, role … but never the section's one-line summary, although E0 produces and persists it (sections rows + pageindex.json)
3. E1 context prefix — implemented
4. ±1 (then ±2) neighbour chunks, same section — PARTIAL: only ±1, no ±2 escalation
5. Known entity hints — MISSING entirely

Today the only consumer of the E0 summaries is the P3 corpusfs projection (root summary per file); every non-root summary is write-only.

Consequences: the design's stated reasons for element 2 ("makes intro/conclusion and list-item-without-preamble decidable") and element 5 (pronoun/partial-name resolution with permission-not-invention semantics) are not available to Selection/Claimify. Note the grounding gate consequence: adding bundle elements means extending the added_context tag enum (header|neighbour|prefix) and _bundle_element, plus an extractor version bump.

Sequencing suggestion: hold until #161 (Claimify-loss ledgering) shows whether context scarcity is actually a measured loss cause — the current funnel evidence points at the verbatim gates, not missing context.

[fazpu]

The section summary element of this issue is now owned by D79 (design PR #164, implementation tracker above): summaries become a verified added_context bundle element (target + ancestors) once the bottom-up summary route lands. This issue narrows to the remaining divergences: known entity hints (explicitly deferred — gated on #148 entity lint plus #161 evidence, see D79 rejected alternatives) and the ±2 neighbour escalation.

[fazpu]

Correction to my previous comment after the PR #164 review cycle: the summary element enters the E2 bundle as orientation context only — there will be no summary added_context kind and no grounding-gate verification against summary text, because both reviewers confirmed that would open a fact-injection channel past the membership-only D32 layer-2 check (an invented phrase in an abstractive summary could be laundered into a claim). Additions must keep coming from source-derived elements (header|neighbour|prefix). See D79 as revised.