[refactor] education.ts hand-rolls entry segmentation instead of using parseEntryBlocks — the repo's #1 complexity function

[s-annam]

Problem

extractEducation is the highest-cognitive-complexity function in the repository, and the reason is that education.ts hand-rolls the entry segmentation that every other extractor gets from the shared parseEntryBlocks primitive.

fallow health (repo-wide, 14,244 functions analyzed) ranks it #1 and #5:

rank	function	cognitive
1	extractEducation (src/lib/heuristics/extract/education.ts:1123)	80
5	educationFromChunk (src/lib/heuristics/extract/education.ts:921)	42

src/lib/heuristics/entry-blocks.ts records the gap in its own prose:

education.ts does not use this primitive at all.
— src/lib/heuristics/entry-blocks.ts:279-280

The three other entry-shaped sections are thin callers of the shared primitive:

• src/lib/heuristics/extract/experience.ts:44 (and a second pass at :58)
• src/lib/heuristics/extract/projects.ts:39
• src/lib/heuristics/extract/achievements.ts:75

education.ts instead imports only two leaf predicates from it — isDateOnlyLine, isEntryHeaderShape (education.ts:22) — and re-implements the rest as a private "chunker": educationFromChunk (:921) joins a chunk with " | " and re-parses fields out of the joined string (:925), with the chunking gates spread across isRealEntryHeader (:258), isCourseworkContinuation (:290), isInlineDatedProgram (:362), isProgramLeadAt (:484) and isInstitutionLeadAt (:545).

The file carries 27 distinct issue references — tied for second-highest in the repo — which is what that hand-rolled path costs in accreted special cases.

Why this is the right shape

parseEntryBlocks is already a parameterized segmentation engine, not an experience-specific one. EntryBlockConfig (entry-blocks.ts:185-213) exposes exactly the axes education needs:

• anchor — education entries anchor on an institution or degree line, not necessarily a date range
• collectBody — education entries may be header-only
• headerLookback — the Degree\nInstitution <dates> stack
• dateParsing: "all" | "date_anchors_only" — education's dates ride the institution line

Projects and achievements were both onboarded onto it by supplying a config. Education is the one holdout.

Scope

Behaviour-preserving segmentation swap. education.ts keeps ownership of its education-specific field logic (degree/field parsing, GPA and honors via education-grade.ts, institution/location splitting, date precision) — the goal is to delete the private chunker, not the domain knowledge.

Expect this to need a new EntryAnchor variant or a widened anchor config; adding one is in the spirit of the primitive and benefits the other three callers.

Implementation plan

1. Characterize the current chunker: what line sequences does it group, and on which fixtures? The education corpus fixtures plus the batch:parser-education lane are the reference set.
2. Express that grouping as an EntryBlockConfig, extending EntryAnchor if the institution-lead / degree-lead cues have no existing anchor.
3. Replace educationFromChunk's chunk-assembly with parseEntryBlocks output, keeping the per-field extraction that runs on each block.
4. Delete the now-dead chunking gates, or demote any that survive as genuine field-level predicates.
5. Re-run fallow health and record the new cognitive numbers for both functions.

Acceptance criteria

• education.ts imports and calls parseEntryBlocks; the private chunk-assembly path is gone.
• Corpus snapshots are byte-identical, or every diff is reviewed fixture-by-fixture and justified in the PR description.
• extractEducation and educationFromChunk both drop below the fallow complexity threshold; the PR quotes before/after.
• npm run verify green; no new KNOWN_FAILURES entries.
• Any new EntryBlockConfig axis is documented in the interface docblock the way the existing four are.

Context

Surfaced by an architecture review of the "should we introduce a plugin architecture?" question. The answer was no — but this convergence was the one structural item the review found that epic #646 does not already cover, and fallow ranks it as the single highest-complexity function in the tree. Related: #646, #831 (batch:parser-education).