PDF detection — Phase 3: remaining four workers (commenting, assessing, linking, tagging)

[adampingel]

Part of #351. Fan the PDF branch out to commenting, assessing, linking, tagging. The geometry tail (buildPdfAnnotation) is shared; each worker's body is already built by its existing processXJob. Shippable one worker at a time.

Per-worker notes

• commenting / assessing — generated-text body; geometry identical to highlighting; lowest risk.
• linking — entity-reference body; occurrence disambiguation matters most here (a wrong occurrence draws a rect on the wrong page); the first-of-many / prefix-suffix case carries more weight.
• tagging — structural; v1 uses the same word-union geometry; block-level bbox is a later refinement, not v1.

Tests (gate this phase)

• Unit/integration: one suite per motivation (mirroring the existing parser tests) — selector geometry + body shape + dedupe.
• e2e: one spec per motivation (or a representative subset — a comment + a linking reference, the two body shapes) on the seeded PDF.
• Live: run each motivation on a real text PDF; confirm geometry + body quality.

Done

All five motivations detect-and-render on text PDFs; per-motivation tests green; live spot-checks.