fix(heuristics): coursework section + multiline-bullet + column deinterleave

[s-annam]

Summary

Fixes the full coursework/multiline-bullet parse path on multi-column student résumés (reproducer: google-docs-skia-proxy-multiline-bullets-coursework.pdf). Three ordered fixes that compose end-to-end:

• "Relevant Coursework" (unrecognized header) is swallowed into the previous entry's description instead of terminating the section #163 — section boundary. "Relevant Coursework"/"Coursework" now opens an education-type section (added aliases + anchor); isVisualHeader gains an ALL-CAPS, font-metadata-independent text-pattern fallback; VISUAL_HEADER_FONT_RATIO lowered 1.2→1.15. Stops unknown headers swallowing the prior entry's description.
• Multi-line (wrapped) bullets truncated at the wrap point in the per-bullet pool, then mis-grouped as "Other" #162 — bullet wrap merge. New mergeWrappedContinuations() in entry-blocks.ts, run per-section in toSectionedResume before the byName flatten; folds wrapped-bullet continuation lines onto their parent so the bullet pool and reconstructed-role description agree — no more truncation or "Other" mis-grouping.
• Multi-column coursework block interleaves across columns during extraction #164 — column de-interleave. New reorderEmbeddedColumns pass upstream of line grouping; de-interleaves localized multi-column blocks (e.g. the coursework grid) to column-major reading order before sectionizing.

Pipeline order is load-bearing: #164 de-interleaves columns → #163's boundary captures the block as education → #162 folds each column's wrap onto its true parent.

Pre-existing Docent tri-line-header mis-parse (bullets in "Other") is orthogonal and left untouched.

Closes #163
Closes #162
Closes #164

Test plan

• npm run typecheck clean
• npm run test green (756 passed, 54 files)
• npm run lint (eslint) clean
• Corpus snapshots regenerated where scoring legitimately shifted (junk coursework bullets removed; specificity/structure rise); no section/field/count loss — diffs eyeballed

[Vaishnavi1709]

Nice composition — the three sub-fixes (#163/#162/#164) read cleanly in the order claimed, and the helper extraction (groupItemsIntoRows, reorderColumnRun) resolves the fallow complexity flag without changing behavior. Tests are well-rationalized; the "does NOT promote Title-Case job title" pin on isTextPatternHeader and the single-row date-rail pin on reorderEmbeddedColumns are exactly the right invariants to lock down.

One must-address before merge, plus one thing worth a sanity check.

🔴 Must address — cascade regression on student-projects-activities-singlecol.expected.json

That snapshot delta isn't only a score uplift — it flips:

cascade.confidence:           0.88 → 0
cascade.suggestedEscalation:  "none" → "ocr"

…which contradicts the PR body's "no section/field/count loss." Reproduced on both branches:

branch	rawChars	extractedChars	ratio	confidence	escalation
main	2279	1288	0.565	0.88	"none"
this PR	2279	1086	0.476	0	"ocr"

Root cause: the bleed you correctly removed used to count toward experience.description (its char sum drops 819 → 617 — exactly the 202-char delta). countExtractedChars (cascade.ts:497) only sums typed fields, so the now-correctly-terminated coursework block — which lands in sections.byName rather than a typed schema field — isn't counted. The extracted/raw ratio drops below EXTRACTION_RATIO_FLOOR = 0.5, collectHardFailures raises low_extraction_ratio, and chooseEscalation routes to "ocr" (confidence.ts:78–82, :148–153, :292).

User-visible score got better (58 → 72), but the cascade now reports this resume as failed and would escalate to OCR — the opposite of the intent.

The other shifted fixtures (weasyprint-cairo-two-column, two-column-achievements-sidebar, awesome-cv-{cv,resume}) all keep their pre-PR cascade.confidence, so the regression is isolated. But the root cause is general: every future "correctly terminated unknown section" you ship has the same risk of dragging the extraction ratio toward the floor.

Pick one:

1. Recommended: extend countExtractedChars to also tally chars from sections.byName beyond the typed-field set — small, scoped, no API change, and protects future fixes too.
2. Have the coursework block populate education[] (e.g. a comma-joined course list as degree) so its content lands in an already-counted field.
3. Acknowledge the regression in the PR body + a comment on the fixture and open a follow-up issue — only if you'd rather not expand scope here.

🟡 Worth a quick look — awesome-cv structure score declined

awesome-cv-cv:     structure 20 → 18, goodBullets 35 → 32 (totalBullets 52)
awesome-cv-resume: structure 23 → 20, goodBullets 23 → 20 (totalBullets 30)

Same total bullets, fewer "good" ones. Specificity ↑ (metricBullets ↑) is expected — wrap-merge revealing metrics in the tail. But the goodBullets drop suggests merged longer bullets are tripping a length cap. Can you confirm this is the expected trade (longer merged bullets correctly reclassify as too-verbose) rather than a downstream rule that should be retuned? Not blocking.

Nits (non-blocking)

• sections.ts ↔ entry-blocks.ts is now a type-only circular import; runtime-safe today, but a value import on the entry-blocks side would make it real. Colocating mergeWrappedContinuations next to toSectionedResume in sections.ts (its only non-test caller) would preempt this.
• isVisualHeader's split && reads oddly — an early return on the font path would mirror the text-pattern path.
• mergeWrappedContinuations does `${prev.text} ${line.text.trim()}` without trimming prev.text — .trimEnd() on prev would avoid double-space on the off chance a line has a trailing space.

Verdict: request changes on the cascade regression; everything else is approve-modulo-nits.

[s-annam]

Thanks for the thorough review, @Vaishnavi1709 — all addressed in 4717423.

🔴 Cascade regression — fixed (option 1)

Extended countExtractedChars (cascade.ts) to also tally sections.byName entries outside the typed-field set (projects/certifications/achievements/other), so a correctly-terminated unknown section no longer drags the extracted/raw ratio toward EXTRACTION_RATIO_FLOOR. Your recommended option — small, scoped, protects future fixes.

student-projects-activities-singlecol is restored to exactly the pre-PR cascade state:

	confidence	escalation
main	0.88	none
this PR (before)	0	ocr
this PR (after fix)	0.88	none

Your "root cause is general" call was right — four other fixtures were already 0/ocr on main for the same reason, and the fix clears all of them too:

• google-docs-skia-proxy-multiline-bullets-coursework: 0/ocr → 0.87/none
• google-docs-skia-proxy-additional-skills: 0/ocr → 0.74/llm
• deedy-resume-macfonts / deedy-resume-openfonts: 0/ocr → 0.64/ner

Score blocks are unchanged across all five — extractedChars feeds confidence/escalation, not the score dimensions. Snapshots rebaked; npm run typecheck + npm run test (756) green.

🟡 awesome-cv goodBullets — confirmed expected

These fixtures aren't in the changed set, so behavior is identical to before this comment. The drop is the wrap-merge trade you predicted: merged longer bullets correctly reclassify against the length cap (verbosity), while specificity rises as the merge reveals tail metrics. Not a downstream rule that needs retuning.

Nits

• mergeWrappedContinuations double-space — fixed: now ${prev.text.trimEnd()} ${line.text.trim()}.
• isVisualHeader && — it already early-returns on the font path (if (isHeaderShort && font≥ratio) return true), so I left it; the only oddness is the wrapped condition, which is cosmetic.
• sections.ts ↔ entry-blocks.ts type-only circular import — deferred. Runtime-safe today (type-only), and colocating mergeWrappedContinuations into sections.ts is a structural move I'd rather not fold into this fix PR. Worth a follow-up if we want to harden it.

Re-requesting review.