refactor(#133): pool anon scorer bullets from accomplishment sections

[s-annam]

Summary

Sources the anonymous scorer's experience-bullet pool from the accomplishment sections (experience / projects / achievements) of SectionedResume via a new extractBulletsFromSections, instead of walking all of rawText and subtracting a skills-line set. Skills are now excluded by construction (skills is not an accomplishment section), retiring extractBulletsFromText, buildSkillsExclusion, and stripBulletMarker — the "pool everything, subtract" shape that forced the #30 side-channel. The experience-completeness check now asks "is there a non-empty experience section?" rather than guessing from the global bullet count (spike §1.5). The #82 live bullet-edit path is threaded so an edit re-grades Specificity/Structure: applyOverrides now takes and returns the edited SectionedResume (new replaceBulletInSections), and App.tsx feeds the edited sections into the scorer.

This is a deliberate scoring change: ATS_SCORE_ALGO_VERSION bumped 1.2 → 1.3 with a changelog entry.

Stacked on #138 (refactor-137-132-fallow-sectioned, the SectionedResume proof PR). Based against that branch, not main — GitHub will auto-retarget to main once #138 merges. Review/merge #138 first.

Resolves #133
Refs #109, #127, #82

Goldens movement (regenerated + spot-audited fixture-by-fixture)

19 of 24 fixtures are version-stamp-only (their bullets were already inside accomplishment sections — unchanged, as expected). 5 moved, all narrowing (totalBullets dropped; none increased → no double-counting):

fixture	totalBullets	overall	why
google-docs-skia-proxy-two-column	18→13	63→74	5 bulleted CONTACT lines in the profile sidebar (• jane@…, • (312)…) were falsely pooled as experience bullets; removed → ratios rise
weasyprint-cairo-two-column	18→13	53→60	same 5 bulleted-contact-sidebar lines leave the pool
awesome-cv-cv	59→52	58→60	6 certifications + 1 education bullet leave the pool
awesome-cv-resume	31→30	67→68	1 education bullet leaves the pool
openresume-react-pdf	12→8	84→84	3 education bullets + 1 project bullet leave the pool (the project bullet lands under a mis-segmented singular "PROJECT" header → education; #109 territory, pre-existing)

Only lossy keys/counts *.expected.json snapshots changed — no fixture binaries touched, so no PII surface.

Test plan

• npm run typecheck clean
• npm run test green (616 tests, 41 files; 24 corpus fixtures match regenerated goldens)
• npm run lint (eslint) clean

[rohithgollapalli]

Review — Approve ✅ (stacked on #138 — merge that first)

A well-scoped, well-documented deliberate scoring change. It removes the "pool all of rawText, then subtract a skills set" anti-pattern and instead pools bullets from the accomplishment sections, making skills exclusion structural rather than a side-channel. The #82 live-edit path is correctly re-threaded through the section view. Verified locally on gh-133.

Verification performed

• ✅ npm run typecheck clean
• ✅ npm run lint clean
• ✅ npm run test — 616/616 pass, 41 files (24 corpus fixtures match the regenerated goldens)
• ✅ No orphaned references to the retired extractBulletsFromText / buildSkillsExclusion / stripBulletMarker — remaining hits are only in docs/segmentation-spike.md
• ✅ App.tsx useMemo deps are complete (state covers state.result.sections; edit overrides listed)

What I checked closely

• Skills-by-construction. extractBulletsFromSections walks only accomplishmentSections (experience/projects/achievements) in policy order; skills can never enter the pool, so the subtraction step and its side-channel are genuinely gone. The reworked unit tests pin this by showing the same lines pool when in experience and don't when in skills.
• Edit → re-lint: make parse overrides authoritative (or honest) #82 edit path. replaceBulletInSections mirrors replaceBulletInRawText (first normalized-match wins, marker preserved), is immutable (clones the map + only the mutated section's array, leaves the input untouched — explicitly asserted in the test), and App.tsx now feeds the edited sections into re-grading. Matching is by normalized text, so the bullet-index reorder (now section order, not rawText order) doesn't affect override application.
• Goldens. All 5 non-stamp movements are narrowing (totalBullets drop, none rise → no double-counting), and the audit table's explanations check out (bulleted contact-sidebar lines and education/cert bullets leaving the pool).

Notes (non-blocking)

1. New invariant coupling in the lone-bullet merge. extractBulletsFromLines dropped the blank-skipping loop and now does j = i + 1, relying on "section lines are already trimmed & non-empty" from toSectionedResume. That holds for both producers today (toSectionedResume and replaceBulletInSections, since empty edits are skipped), and it's documented — but it's a real coupling: any future code that puts a blank line into a byName array would make a lone glyph adopt "" as its text. Worth keeping in mind if the section view gains other producers.
2. Cross-section lone glyph. A marker-only line stranded at the end of a section no longer merges with text in the following section (it breaks within the section). Negligible and arguably more correct, just flagging the semantic shift from the whole-document walk.
3. openresume-react-pdf drops a genuine project bullet due to a mis-segmented singular "PROJECT" header landing under education (Generalize resume section recognition (layered recognizer) #109, pre-existing). The PR rightly calls this out — but it's worth noting the new "pool by section membership" model makes pre-existing segmentation bugs score-visible in a way the old rawText walk masked. Raises the value of Generalize resume section recognition (layered recognizer) #109.
4. Experience-completeness predicate has no corpus coverage for its changed branch. passed: expEntries.length > 0 || (experience section non-empty) replaces the old || bullets.length > 0. In all 5 moved fixtures completeness stayed at 27, so the divergent case (bullets only in projects/achievements, no experience section) is exercised only by the flags missing sections unit test, not the corpus. Fine, just no fixture pins it.

None block merge. Clean structural win.

🤖 Generated with Claude Code

The base branch was changed.

[s-annam]

Rebased onto main now that #138 is merged. The PR was auto-closed when its stacked base branch (refactor-137-132-fallow-sectioned) was deleted on #138's squash-merge; reopened and retargeted to main. Since #138 landed as a squash commit, this branch's copy of the #137/#132 work conflicted — dropped it via git rebase --onto origin/main d09fca2, leaving a single clean commit (a466107) on top of main.

Gates green on the rebased branch: npm run typecheck clean, npm run lint clean, npm run test 616/616. Mergeable is now clean (was CONFLICTING).

On the four non-blocking review notes: all are flagging semantic shifts / pre-existing conditions, not change requests, so leaving code as-is. Note 4 (no corpus fixture pins the changed experience-completeness branch) is the only one with a possible follow-up — the flags missing sections unit test covers it; deferring a dedicated fixture as out of scope for this structural refactor.