INST-5b · Store all three classifications + materialized agreement booleans

> Item derived from the fact-checking cycle analysis (**2026-07-03, v2**). Split from INST-5 (#2495).

### Action
Store the three classifications separately (initial / crossCheck / final) + materialized booleans:
- `agreement.initialVsCross`
- `agreement.crossVsFinal`

### Impact
Inter-rater and adoption become a simple `$avg`. Builds on INST-5a (#2495 → 5a) by adding the cross/final fields and the agreement flags.

### Note (D2)
Do **not** cite the 44.4% "adoption in divergent cases" figure as fact until 5a/5b ship — it depends on an independent `final` that does not exist today. Citable now: only the 84.5% (N=187) initial↔cross concordance.

---
- **Priority:** 🟡 Medium · **Effort:** M
- **Parent:** #2495 · **Epic:** #2489

Start by reading parent issue #2495 and the 2026-07-03 v2 fact-checking cycle analysis referenced here, then trace how the current classifications are stored. Confirm the data path and define done as persistence of initial, crossCheck, and final classifications plus both materialized agreement booleans, with values suitable for the stated aggregate calculations.

Vom Indexierungsmodell aus dem Issue-Text verfasst.