Theorizer plugin output lags web: dense theory statements + prose-only novelty

# Theorizer plugin output lags the web surface: dense theory statements + prose-only novelty

_Sourced from Asta feedback submission `2026-08-12-theorizer-web-viewer` (retrieved from the feedback bucket). The submitter has visibility into the Asta web product and is comparing surfaces directly._

## Skills

`generate-theories` (theory-description rendering) and its novelty-evaluation output (`evaluate-novelty`).

## Current behavior

Improvements made on the Asta **web** surface have not propagated to the plugin/CLI surface:

- Theories arrive as dense `FORK:` statements with stacked numeric thresholds — the web version reworked these to plainer language.
- Novelty output is repetitive per-dimension prose ("What is Known / What is Introduced / What is Novel" blocks repeated per statement), with **no scales, no structured labels, and no per-claim verdict fields**. A downstream agent had to grep prose to synthesize verdicts. The web version presents novelty dimensions/scales in a structured, less-jargon form.

The two surfaces are drifting apart.

## Desired behavior

- Port the web version's plainer theory-description style into the plugin output.
- Emit **structured novelty verdicts** (per-claim labels/scores, not just prose) so downstream agents can consume them without parsing text.
- Ideally, both surfaces render from one shared presentation layer so future web-side improvements don't require a second manual port.

## Related

- CLI-shape doc mismatch for `generate-theories`: already tracked in #114.
- Silent local-PDF provenance gap: filed separately (Theorizer PDF-provenance issue).

Suggested-by: @jbragg