Shared lexicon (a): one bullet-glyph vocabulary — ⁃ and — are parser bullets the scorer silently drops

[s-annam]

Part (a) of #653. Not a good first issue — this one changes scoring behaviour and needs per-glyph human adjudication.

Problem

The bullet-glyph vocabulary exists as divergent copies, and the divergence is live: glyphs the parser treats as bullets are invisible to the scorer, so those bullets drop out of the Specificity pool.

Verified divergence:

site	class
src/lib/heuristics/line-primitives.ts:35 (BULLET_CLASS)	[•‣▪●◦⁃*\-–—]
src/lib/heuristics/sections.ts:145 (VISUAL_BULLET_RE)	[•‣▪●◦⁃*\-–—]
src/lib/heuristics/regex.ts:321 (LEADING_BULLET_RE)	[•‣▪●◦⁃*\-–—]
src/lib/score/score.ts:136 (BULLET_MARKER_RE)	[-*•●–▪◦‣▶►·�]
src/lib/score/score.ts:789 (LONE_BULLET_RE)	[•●▪◦‣▶►·�]
src/lib/edit/apply-overrides.ts:105 (LEADING_MARKER_RE)	[-*•●–▪◦‣▶►·�]

Plus copies in markdown-emit.ts, extract/group-bullets.ts and webllm/localize/achievements.ts — 8 in total per the #646 review.

⁃ (U+2043 hyphen bullet) and — (em dash) are parser bullets absent from the scorer's class. A résumé using either has its Experience bullets segmented correctly by the parser and then silently skipped by extractBulletsFromLines (score.ts:808-822), deflating Specificity. Conversely ▶, ►, · and � are scorer-only.

Why this is separate from #653(b)/(c)/(d)

Merging the sets changes scores on existing résumés. That is a deliberate behaviour change requiring a version bump and a reviewed rebaseline, which is why it cannot ride along with the behaviour-preserving consolidations.

Implementation plan

1. Mint a new import-nothing leaf for the glyph vocabulary. House precedent is extract/title-shape.ts. Not line-primitives.ts — it imports regex.ts, and deriving there re-creates the cycle regex.ts:317-321 exists to avoid, dragging the 672-line regex.ts and its module-eval Intl gazetteer onto the eager entry graph via the score/edit consumers (the feat: primary role targeting, editable headline, and job-search handoff (#598) #605 failure class).
2. Adjudicate each glyph explicitly: for every symbol in the union, decide parser-bullet / scorer-bullet / both, and write the reason down. - and – in particular are ambiguous (a leading hyphen is also a date-range fragment and a negated number — see cleanRewriteLine eats the minus sign off a line-initial negative number (-5% churn → 5% churn) #821).
3. Derive all 8 copies from the leaf as named exported sets, preserving any distinction that survives adjudication as its own derived constant rather than collapsing it.
4. Bump ATS_SCORE_ALGO_VERSION.
5. Rebaseline the 54 corpus snapshots; review the score deltas fixture by fixture.

Acceptance criteria

• One import-nothing leaf owns the glyph vocabulary; all 8 sites derive from it.
• The leaf imports nothing, asserted by a test (follow title-shape.ts's contract).
• Per-glyph adjudication is written down in the leaf's docblock, not just in the PR.
• ATS_SCORE_ALGO_VERSION bumped.
• Corpus rebaseline reviewed fixture-by-fixture; score deltas explained in the PR description.
• A regression test covers a ⁃-bulleted and an —-bulleted résumé reaching the Specificity pool.
• The / entry chunk does not gain regex.ts (build-verified).

Parent: #653. Roadmap item 5(a) of #646.